Sei sulla pagina 1di 173

c 

Forecasting Conflict in the Balkans Peramalan Konflik di Balkan


using Hidden Markov Models menggunakan Hidden Markov Model
Philip A. Schrodt Philip A. Schrodt
Department of Political Science Departemen Ilmu Politik
University of Kansas University of Kansas
Lawrence, KS 66045 USA Lawrence, KS 66045 USA
phone: +1.785.864.9024 fax: +1.785.864.5700 Telepon: +1.785.864.9024 fax: +1.785.864.5700
p-schrodt@ukans.edu p-schrodt@ukans.edu
August 2000 Agustus 2000
Paper presented at the American Political Science Association meetings Makalah disajikan pada
pertemuan Ilmu Politik American Association
Washington, DC Washington, DC
Electronic copies of this paper are available at http://www.ukans.edu/keds/papers.html and at
salinan elektronik dari tulisan ini tersedia di http://www.ukans.edu/keds/papers.html dan di
APSA PROceedings webs site: http://PRO.harvard.edu. APSA Prosiding web site:
http://PRO.harvard.edu. This work was partially supported by an Karya ini sebagian didukung
oleh
EPSCoR start-up grant from the National Computational Science Alliance and utilized the
NCSA EPSCoR start-up hibah dari National Computational Science Alliance dan memanfaatkan
NCSA
SGI/CRAY Origin2000 parallel computing system. SGI / Cray Origin2000 sistem komputasi
paralel
c 
Forecasting Conflict in the Balkans Peramalan Konflik di Balkan
using Hidden Markov Models menggunakan Hidden Markov Model
Philip A. Schrodt Philip A. Schrodt
Department of Political Science Departemen Ilmu Politik
University of Kansas University of Kansas
Lawrence, KS 66045 USA Lawrence, KS 66045 USA
phone: +1.785.864.9024 fax: +1.785.864.5700 Telepon: +1.785.864.9024 fax: +1.785.864.5700
p-schrodt@ukans.edu p-schrodt@ukans.edu
August 2000 Agustus 2000
Paper presented at the American Political Science Association meetings Makalah disajikan pada
pertemuan Ilmu Politik American Association
Washington, DC Washington, DC
Electronic copies of this paper are available at http://www.ukans.edu/keds/papers.html and at
salinan elektronik dari tulisan ini tersedia di http://www.ukans.edu/keds/papers.html dan di
APSA PROceedings webs site: http://PRO.harvard.edu. APSA Prosiding web site:
http://PRO.harvard.edu. This work was partially supported by an Karya ini sebagian didukung
oleh
EPSCoR start-up grant from the National Computational Science Alliance and utilized the
NCSA EPSCoR start-up hibah dari National Computational Science Alliance dan memanfaatkan
NCSA
SGI/CRAY Origin2000 parallel computing system. SGI / Cray Origin2000 sistem komputasi
paralel.
c  c 
Abstract Abstrak
This study uses hidden Markov models (HMM) to forecast conflict in the former Yugoslavia for
Penelitian ini menggunakan model Markov tersembunyi (HMM) untuk meramalkan konflik di
bekas Yugoslavia untuk
the period January 1991 through January 1999. periode Januari 1991 hingga Januari 1999. The
political and military events reported in the Dan militer peristiwa politik yang dilaporkan dalam
lead sentences of Reuters news service stories were coded into the World Events Interaction
Survey kalimat utama cerita layanan berita Reuters diberi kode ke dalam Peristiwa Interaksi
Dunia Survey
(WEIS) event data scheme. (Weis) acara skema data. The forecasting scheme involved randomly
selecting eight 100-event Skema peramalan yang terlibat secara acak memilih delapan 100-acara
"templates" taken at a 1-, 3- or 6-month forecasting lag for high-conflict and low-conflict weeks.
"Template" diambil pada 1 -, 3 - atau 6 bulan lag peramalan untuk tinggi-konflik dan konflik
minggu rendah. A A
separate HMM is developed for the high-conflict-week sequences and the low-conflict-week
HMM terpisah dikembangkan untuk konflik-minggu urutan tinggi dan rendah konflik-minggu
sequences. urutan. Forecasting is done by determining whether a sequence of observed events fit
the high- Peramalan dilakukan dengan menentukan apakah suatu urutan peristiwa yang diamati
sesuai dengan tinggi
conflict or low-conflict model with higher probability. konflik-konflik atau model rendah dengan
probabilitas yang lebih tinggi.
Models were selected to maximize the difference between correct and incorrect predictions,
Model dipilih untuk memaksimalkan perbedaan antara benar dan salah prediksi,
evaluated by week. dievaluasi oleh minggu. Three weighting schemes were used: unweighted
(U), penalize false positives Tiga skema pembobotan digunakan: unweighted (U), menghukum
positif palsu
(P) and penalize false negatives (N). (P) dan menghukum palsu negatif (N). There is a relatively
high level of convergence in the Ada yang relatif tinggi tingkat konvergensi dalam
estimates²the best and worst models of a given type vary in accuracy by only about 15% to
20%. perkiraan-dan terburuk model terbaik dari jenis yang diberikan bervariasi dalam akurasi
hanya 15% sekitar sampai 20%.
In full-sample tests, the U and P models produce at overall accuracy of around 80%. Di-sampel
tes penuh, U dan model P menghasilkan dengan ketelitian keseluruhan sekitar 80%. However,
Namun,
these models correctly forecast only about 25% of the high-conflict weeks, although about 60%
of model ini benar perkiraan hanya sekitar 25% dari konflik minggu tinggi, meskipun sekitar
60% dari
the cases where a high-conflict week has been forecast turn out to have high conflict. kasus
dimana konflik minggu tinggi telah diperkirakan ternyata memiliki konflik yang tinggi. In
contrast, Sebaliknya,
the N model has an overall accuracy of only about 50% in full-sample tests, but it correctly
model N memiliki akurasi keseluruhan hanya sekitar 50% di-sampel tes penuh, tapi dengan
benar
forecasts high-conflict weeks with 85% accuracy in the 3- and 6-month horizon and 92%
accuracy ramalan tinggi-konflik minggu dengan akurasi 85% dalam 3 - dan-bulan cakrawala 6
dan akurasi 92%
in the 1-month horizon. di bulan cakrawala 1. However, this is achieved by excessive predictions
of high-conflict weeks: Namun, hal ini dicapai dengan prediksi yang berlebihan-konflik minggu
tinggi:
only about 30% of the cases where a high-conflict week has been forecast are high-conflict.
hanya sekitar 30% dari kasus-kasus dimana-konflik minggu tinggi telah memperkirakan tinggi-
konflik.
Models that use templates from only the previous year usually do about as well as models based
on Model yang menggunakan template hanya dari tahun sebelumnya biasanya sekitar serta
model berdasarkan
the entire sample. seluruh sampel.
The models are remarkably insensitive to the length of the forecasting horizon²the drop-off in
Model ini sangat sensitif terhadap panjang peramalan cakrawala-drop-off di
accuracy at longer forecasting horizons is very small, typically around 2%-4%. akurasi di
cakrawala peramalan lagi sangat kecil, biasanya sekitar 2% -4%. There is also no Juga tidak ada
clear difference in the estimated coefficients for the 1-month and 6-month models. jelas
perbedaan koefisien estimasi untuk bulan 1 dan model 6 bulan. An extensive Sebuah luas
analysis was done of the coefficient estimates in the full-sample model to determine what the
model Analisis dilakukan terhadap estimasi koefisien dalam model-sampel penuh untuk
menentukan model
was "looking at" in order to make predictions. adalah "melihat" untuk membuat prediksi. While a
number of statistically significant Sementara sejumlah statistik signifikan
differences exist between the high and low conflict models, these do not fall into any neat
patterns. ada perbedaan antara rendah konflik model dan tinggi, ini tidak jatuh ke dalam suatu
pola rapi.
This is probably due to a combination of the large number of parameters being estimated, the Hal
ini mungkin karena kombinasi dari sejumlah besar parameter yang diestimasi,
multiple local maxima in the estimation surface, and the complications introduced by the
presence beberapa lokal maxima di permukaan estimasi, dan komplikasi diperkenalkan oleh
kehadiran
of a number of very low probability event categories. sejumlah peristiwa probabilitas rendah
kategori sangat. Some experiments with simplified models Beberapa percobaan dengan model
yang disederhanakan
indicate that it is possible to use models with substantially fewer parameters without markedly
menunjukkan bahwa adalah mungkin untuk menggunakan model dengan sedikit parameter
substansial tanpa nyata
decreasing the accuracy of the predictions; in fact predictions of the high conflict periods
actually mengurangi akurasi prediksi, dalam kenyataan prediksi periode konflik tinggi benar-
benar
increase in accuracy quite substantially. peningkatan akurasi yang cukup substansial

c  c 
Forecasting Conflict in the Balkans Peramalan Konflik di Balkan
using Hidden Markov Models menggunakan Hidden Markov Model
Philip A. Schrodt Philip A. Schrodt
Department of Political Science Departemen Ilmu Politik
University of Kansas University of Kansas
Lawrence, KS 66045 USA Lawrence, KS 66045 USA
phone: +1.785.864.9024 fax: +1.785.864.5700 Telepon: +1.785.864.9024 fax: +1.785.864.5700
p-schrodt@ukans.edu p-schrodt@ukans.edu
August 2000 Agustus 2000
Paper presented at the American Political Science Association meetings Makalah disajikan pada
pertemuan Ilmu Politik American Association
Washington, DC Washington, DC
Electronic copies of this paper are available at http://www.ukans.edu/keds/papers.html and at
salinan elektronik dari tulisan ini tersedia di http://www.ukans.edu/keds/papers.html dan di
APSA PROceedings webs site: http://PRO.harvard.edu. APSA Prosiding web site:
http://PRO.harvard.edu. This work was partially supported by an Karya ini sebagian didukung
oleh
EPSCoR start-up grant from the National Computational Science Alliance and utilized the
NCSA EPSCoR start-up hibah dari National Computational Science Alliance dan memanfaatkan
NCSA
SGI/CRAY Origin2000 parallel computing system. SGI / Cray Origin2000 sistem komputasi
paralel.

c  c 
Abstract Abstrak
This study uses hidden Markov models (HMM) to forecast conflict in the former Yugoslavia for
Penelitian ini menggunakan model Markov tersembunyi (HMM) untuk meramalkan konflik di
bekas Yugoslavia untuk
the period January 1991 through January 1999. periode Januari 1991 hingga Januari 1999. The
political and military events reported in the Dan militer peristiwa politik yang dilaporkan dalam
lead sentences of Reuters news service stories were coded into the World Events Interaction
Survey kalimat utama cerita layanan berita Reuters diberi kode ke dalam Peristiwa Interaksi
Dunia Survey
(WEIS) event data scheme. (Weis) acara skema data. The forecasting scheme involved randomly
selecting eight 100-event Skema peramalan yang terlibat secara acak memilih delapan 100-acara
"templates" taken at a 1-, 3- or 6-month forecasting lag for high-conflict and low-conflict weeks.
"Template" diambil pada 1 -, 3 - atau 6 bulan lag peramalan untuk tinggi-konflik dan konflik
minggu rendah. A A
separate HMM is developed for the high-conflict-week sequences and the low-conflict-week
HMM terpisah dikembangkan untuk konflik-minggu urutan tinggi dan rendah konflik-minggu
sequences. urutan. Forecasting is done by determining whether a sequence of observed events fit
the high- Peramalan dilakukan dengan menentukan apakah suatu urutan peristiwa yang diamati
sesuai dengan tinggi
conflict or low-conflict model with higher probability. konflik-konflik atau model rendah dengan
probabilitas yang lebih tinggi.
Models were selected to maximize the difference between correct and incorrect predictions,
Model dipilih untuk memaksimalkan perbedaan antara benar dan salah prediksi,
evaluated by week. dievaluasi oleh minggu. Three weighting schemes were used: unweighted
(U), penalize false positives Tiga skema pembobotan digunakan: unweighted (U), menghukum
positif palsu
(P) and penalize false negatives (N). (P) dan menghukum palsu negatif (N). There is a relatively
high level of convergence in the Ada yang relatif tinggi tingkat konvergensi dalam
estimates²the best and worst models of a given type vary in accuracy by only about 15% to
20%. perkiraan-dan terburuk model terbaik dari jenis yang diberikan bervariasi dalam akurasi
hanya 15% sekitar sampai 20%.
In full-sample tests, the U and P models produce at overall accuracy of around 80%. Di-sampel
tes penuh, U dan model P menghasilkan dengan ketelitian keseluruhan sekitar 80%. However,
Namun,
these models correctly forecast only about 25% of the high-conflict weeks, although about 60%
of model ini benar perkiraan hanya sekitar 25% dari konflik minggu tinggi, meskipun sekitar
60% dari
the cases where a high-conflict week has been forecast turn out to have high conflict. kasus
dimana konflik minggu tinggi telah diperkirakan ternyata memiliki konflik yang tinggi. In
contrast, Sebaliknya,
the N model has an overall accuracy of only about 50% in full-sample tests, but it correctly
model N memiliki akurasi keseluruhan hanya sekitar 50% di-sampel tes penuh, tapi dengan
benar
forecasts high-conflict weeks with 85% accuracy in the 3- and 6-month horizon and 92%
accuracy ramalan tinggi-konflik minggu dengan akurasi 85% dalam 3 - dan-bulan cakrawala 6
dan akurasi 92%
in the 1-month horizon. di bulan cakrawala 1. However, this is achieved by excessive predictions
of high-conflict weeks: Namun, hal ini dicapai dengan prediksi yang berlebihan-konflik minggu
tinggi:
only about 30% of the cases where a high-conflict week has been forecast are high-conflict.
hanya sekitar 30% dari kasus-kasus dimana-konflik minggu tinggi telah memperkirakan tinggi-
konflik.
Models that use templates from only the previous year usually do about as well as models based
on Model yang menggunakan template hanya dari tahun sebelumnya biasanya sekitar serta
model berdasarkan
the entire sample. seluruh sampel.
The models are remarkably insensitive to the length of the forecasting horizon²the drop-off in
Model ini sangat sensitif terhadap panjang peramalan cakrawala-drop-off di
accuracy at longer forecasting horizons is very small, typically around 2%-4%. akurasi di
cakrawala peramalan lagi sangat kecil, biasanya sekitar 2% -4%. There is also no Juga tidak ada
clear difference in the estimated coefficients for the 1-month and 6-month models. jelas
perbedaan koefisien estimasi untuk bulan 1 dan model 6 bulan. An extensive Sebuah luas
analysis was done of the coefficient estimates in the full-sample model to determine what the
model Analisis dilakukan terhadap estimasi koefisien dalam model-sampel penuh untuk
menentukan model
was "looking at" in order to make predictions. adalah "melihat" untuk membuat prediksi. While a
number of statistically significant Sementara sejumlah statistik signifikan
differences exist between the high and low conflict models, these do not fall into any neat
patterns. ada perbedaan antara rendah konflik model dan tinggi, ini tidak jatuh ke dalam suatu
pola rapi.
This is probably due to a combination of the large number of parameters being estimated, the Hal
ini mungkin karena kombinasi dari sejumlah besar parameter yang diestimasi,
multiple local maxima in the estimation surface, and the complications introduced by the
presence beberapa lokal maxima di permukaan estimasi, dan komplikasi diperkenalkan oleh
kehadiran
of a number of very low probability event categories. sejumlah peristiwa probabilitas rendah
kategori sangat. Some experiments with simplified models Beberapa percobaan dengan model
yang disederhanakan
indicate that it is possible to use models with substantially fewer parameters without markedly
menunjukkan bahwa adalah mungkin untuk menggunakan model dengan sedikit parameter
substansial tanpa nyata
decreasing the accuracy of the predictions; in fact predictions of the high conflict periods
actually mengurangi akurasi prediksi, dalam kenyataan prediksi periode konflik tinggi benar-
benar
increase in accuracy quite substantially. peningkatan akurasi yang cukup substansial.

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 1 Page 1
The Sequence Recognition Approach to Political Pengakuan Sequence Pendekatan Politik
Forecasting Peramalan
Event sequences are a key element in human reasoning about international events. Event sekuens
adalah elemen kunci dalam pemikiran manusia tentang kejadian internasional. Human Manusia
analysts "understand" an international situation when they recognize sequences of political
activity analis "memahami" situasi internasional ketika mereka mengenali urutan kegiatan politik
corresponding to those observed in the past. sesuai dengan yang diamati di masa lalu. Empirical
and anecdotal evidence point to the Anekdot bukti empiris dan menunjuk ke
likelihood that humans have available in long-term associative memory a set of "templates" for
kemungkinan bahwa manusia telah tersedia dalam jangka panjang asosiatif memori satu set
"template" untuk
common sequences of actions that can occur in the international system (and in social situations
umum urutan tindakan yang dapat terjadi dalam sistem internasional (dan dalam situasi sosial
generally). umumnya). When part of a sequence is matched, the analyst predicts that the
remainder of the Ketika bagian dari urutan yang cocok, analis memprediksi bahwa sisa dari
sequence will be carried out m   
though often the analyst will make a prediction for
the urutan akan dilakukan m   
walaupun sering analis akan membuat prediksi untuk
express purpose of insuring that the remainder of the sequence is  carried out.
mengungkapkan tujuan menjamin bahwa sisa urutan   dilakukan. Sequences can Urutan
dapat
be successfully matched by human analysts in the presence of noise and incomplete information,
akan berhasil dicocokkan oleh analis manusia di hadapan kebisingan dan informasi yang tidak
lengkap,
and can also be used to infer events that are not directly observed but which are necessary dan
juga dapat digunakan untuk menyimpulkan peristiwa yang tidak secara langsung diamati tetapi
yang diperlukan
prerequisites for events that have been observed. prasyarat untuk acara yang telah diamati.
The use of analogy or "precedent-based reasoning" has been advocated as a key cognitive
Penggunaan analogi atau "berbasis penalaran preseden" telah dianjurkan sebagai kunci kognitif
mechanism in the analysis of international politics by Alker (1987), Mefford (1985, 1991) and
mekanisme dalam analisis politik internasional dengan Alker (1987), Mefford (1985, 1991) dan
others, and is substantially different from the statistical, dynamic and rational choice paradigms
that lain, dan secara substansial berbeda dari yang dinamis dan rasional pilihan paradigma,
statistik yang
characterize most contemporary quantitative models of international behavior. ciri kuantitatif
model kontemporer paling perilaku internasional. Khong (1992) and Khong (1992) dan
Vertzberger (1990) review the general arguments in the cognitive psychology literature on use of
Vertzberger (1990) meninjau argumen umum dalam literatur psikologi kognitif pada penggunaan
analogy in political reasoning; May (1973) and Neustadt and May (1986) discuss it from a more
analogi dalam penalaran politik; Mei (1973) dan Neustadt dan Mei (1986) membahas dari yang
lebih
pragmatic and policy-oriented perspective. pragmatis dan berorientasi kebijakan perspektif. As
Khong observes: Sebagai Khong mengamati:
Simply stated, ... Secara sederhana, ... analogies are cognitive devices that "help" policymakers
perform six Analogi adalah alat kognitif yang "membantu" para pembuat kebijakan melakukan
enam
diagnostic tasks central to political decision-making. Pembuatan diagnostik tugas untuk pusat
keputusan politik. Analogies (1) help define the nature of Analogi (1) membantu menentukan
sifat
the situation confronting the policymaker; (2) help assess the stakes, and (3) provide situasi yang
dihadapi pembuat kebijakan, (2) membantu menilai taruhannya, dan (3) menyediakan
prescriptions. resep. They help evaluate alternative options by (4) predicting their chances of
success, Mereka membantu mengevaluasi pilihan-pilihan alternatif (4) memprediksi peluang
keberhasilan,
(5) evaluating their moral rightness and (6) warning about the dangers associated with options.
(5) mengevaluasi kebenaran moral mereka dan (6) peringatan tentang bahaya yang berkaitan
dengan opsi.
(pg. 10) (Hal. 10)
The ubiquity of analogical reasoning is supported by a plethora of experimental studies in Di
mana-mana penalaran analogis didukung oleh kebanyakan studi eksperimental di
cognitive psychology in addition to the case studies from the foreign policy literature. kognitif
psikologi di samping studi kasus dari literatur kebijakan luar negeri.
Analogical reasoning is an easy task for the human brain, one that is substantially easier than
penalaran analogis adalah tugas yang mudah bagi otak manusia, yang secara substansial lebih
mudah daripada
sequential or deductive reasoning. berurutan atau penalaran deduktif. Most experimental
evidence suggests that human memory is Sebagian besar bukti eksperimental menunjukkan
bahwa memori manusia
organized so that when one item is recalled, this naturally activates links to other items that have
terorganisir sehingga ketika satu item diingat, ini alami mengaktifkan link ke item lain yang
memiliki
features in common, and these are more likely to be recalled as well (Anderson 1983; Kohonen
fitur yang sama, dan ini lebih cenderung diingat juga (Anderson 1983; Kohonen
1984). 1984).

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 2 Page 2
Because analogies are so prevalent in human political reasoning, it would be helpful to have
Karena analogi-analogi yang begitu umum dalam penalaran politik manusia, itu akan membantu
untuk memiliki
some computational method for systematically assessing the similarity of two sequences of
beberapa metode komputasi untuk secara sistematis menilai kemiripan dari dua urutan
political events. peristiwa politik. In Schrodt (1991), I posed this problem in the following
manner: Dalam Schrodt (1991), saya berpose masalah ini dengan cara berikut:
In human pattern recognition, we have a general idea of what a category of event sequences
Dalam pengenalan pola manusia, kita memiliki gambaran umum tentang apa urutan kategori
acara
look like²the archetypal war, the archetypal coup, and so forth. terlihat seperti-perang pola
dasar, kudeta tipikal, dan sebagainya. In a sense, ideal sequences Dalam arti, ideal urutan
are the centroid of a cluster of sequences, but that centroid is a sequence rather than a point.
adalah centroid dari sekelompok urutan, tetapi centroid itu adalah urutan daripada suatu titik. If
Jika
a method could be found for constructing such a sequence, the cluster of behaviors could be
metode dapat ditemukan untuk membangun seperti berurutan, cluster perilaku dapat
represented by the single ideal sequence, which would substantially reduce computing time and
diwakili oleh urutan yang ideal tunggal, yang secara substansial akan mengurangi waktu
komputasi dan
provide some theoretical insights as to the distinguishing characteristics of a cluster.
menyediakan beberapa wawasan teoritis mengenai karakteristik yang membedakan sebuah
cluster. (pg. 186) (Hal. 186)
The problem of generalizing sequences is particularly salient to the analysis of international
Masalah generalisasi urutan sangat penting untuk analisis internasional
political behavior in the late 20th century because many contemporary situations do not have
exact perilaku politik pada abad ke-20 terlambat karena situasi kontemporer banyak yang tidak
memiliki tepat
historical analogs. sejarah analog. Yet human analysts are clearly capable of making analogies
based on some Namun analis manusia jelas mampu membuat analogi berdasarkan pada beberapa
characteristics of those behaviors. Karakteristik dari perilaku. For example, because of its
unusual historical circumstances, Sebagai contoh, karena keadaan yang tidak biasa historis,
Zaire in 1997 had a number of unique characteristics, but nonetheless analysts pieced together
Zaire pada tahun 1997 memiliki sejumlah karakteristik yang unik, tapi tetap disatukan analis
sufficient similarities between Zaire and a variety of earlier crises in Africa and elsewhere to
come cukup kesamaan antara Zaire dan berbagai krisis sebelumnya di Afrika dan tempat lain
untuk datang
to the correct conclusion that Zaire had entered a period of rapid political change. pada
kesimpulan yang benar bahwa Zaire telah memasuki masa perubahan politik yang cepat. The key
to this Kunci ini
was the ability to use   analogies: if one insisted on matching all of the features of a adalah
kemampuan untuk menggunakan analogi   jika seseorang bersikeras cocok semua fitur dari
case²which a human analyst would almost never do, but a computer might²then the Zairian
kasus yang seorang analis manusia akan hampir tidak pernah dilakukan, tetapi komputer
mungkin-maka Zairian
situation would be nearly impossible to classify using analogies. situasi akan hampir mustahil
untuk mengklasifikasikan menggunakan analogi.
Techniques for comparing two sequences of discrete events²nominal-level variables occurring
Teknik untuk membandingkan dua variabel urutan peristiwa-nominal-level diskrit terjadi
over time²are poorly developed compared to the huge literature involving the study of interval-
dari waktu ke waktu-yang kurang berkembang dibandingkan dengan literatur yang besar yang
melibatkan studi interval-
level time series. tingkat time series. Nonetheless, several methods are available, and the
problem has received Meskipun demikian, beberapa metode yang tersedia, dan masalah telah
menerima
considerable attention in the past three decades because it is important in the problems of
studying perhatian yang cukup besar dalam tiga dekade terakhir karena penting dalam masalah
belajar
genetic sequences in DNA, and computer applications in involving human speech recognition.
urutan genetik dalam DNA, dan aplikasi komputer dalam melibatkan pengenalan suara manusia.
Both of these problems have potentially large economic payoffs, which tends to correlate with
the Kedua masalah ini berpotensi besar ekonomi hadiah, yang cenderung berkorelasi dengan
expenditure of research efforts. pengeluaran upaya penelitian. Until fairly recently, one of the
most common techniques was the Sampai akhir-akhir ini, salah satu teknik yang paling umum
adalah
Levenshtein metric (see Kruskal 1983; Sankoff & Kruskall 1983); Schrodt (1991) uses this in a
Levenshtein metrik (lihat Kruskal 1983; Sankoff & Kruskall 1983); Schrodt (1991)
menggunakan ini dalam sebuah
study of the BCOW crises. studi tentang krisis BCOW. Other non-linear methods such as neural
networks, genetic Non-linear metode seperti jaringan saraf, genetik
algorithms, and locating common subsets within the sequences (Bennett & Schrodt 1987;
Schrodt algoritma, dan subset umum lokasi dalam urutan (Bennett & Schrodt 1987; Schrodt
1990) have also been used. 1990) juga telah digunakan.
Hidden Markov models Model Markov tersembunyi
Over the past decade the hidden Markov model (HMM) has emerged as one of the most Selama
satu dekade terakhir model Markov tersembunyi (HMM) telah muncul sebagai salah satu yang
paling
widely used techniques for the classification of noisy sequences into a set of discrete categories
banyak digunakan teknik untuk klasifikasi urutan bising menjadi seperangkat kategori diskrit
c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 3 Page 3
(or, equivalently, computing the probability that a given sequence was generated by a known
(Atau, sama, komputasi probabilitas bahwa urutan yang diberikan dihasilkan oleh yang dikenal
general model).sequence comparison method. umum model). urutan metode perbandingan.
While the most common applications of HMMs Sedangkan aplikasi yang paling umum dari
HMMs
are found in speech recognition and comparing protein sequences, a recent search of the World
ditemukan dalam pengenalan suara dan urutan protein membandingkan, pencarian baru-baru ini
Dunia
Wide Web found applications in fields as divergent as modeling the control of cellular phone
Wide Web ditemukan aplikasi dalam bidang yang berbeda seperti model kontrol telepon selular
networks, computer recognition of American Sign Language and²inevitably²the timing of
jaringan, komputer pengakuan dari American Sign Language dan-pasti-waktu
trading in financial markets. perdagangan di pasar keuangan. The purpose of this project is to
apply this technique to the problem Tujuan dari proyek ini adalah untuk menerapkan teknik ini
untuk masalah
of forecasting conflict in the former Yugoslavia. peramalan konflik di bekas Yugoslavia.
A sequence is "noisy" when it contains missing, erroneous and extraneous elements, and urutan
adalah "berisik" ketika mengandung yang hilang, dan unsur-unsur asing yang salah, dan
consequently the sequence cannot be classified by simply matching it to a set of known "correct"
urutan akibatnya tidak dapat diklasifikasikan dengan hanya cocok untuk satu set dikenal "benar"
sequences. urutan. A spelling program, for example, would always mark "wan" as an incorrect
spelling of Sebuah program ejaan, misalnya, selalu akan menandai "wan" sebagai salah ejaan
"one" because written English usually allows one and only one correct spelling of a word. "Satu"
karena ditulis bahasa Inggris biasanya memungkinkan satu dan hanya satu ejaan yang benar dari
sebuah kata. Spoken Lisan
English, in contrast, allows a wide variation of pronunciations, and in some regional dialects,
"wan" is Bahasa Inggris, sebaliknya, memungkinkan variasi luas lafal, dan dalam beberapa
dialek regional, "wan" adalah
the most common pronunciation of "one". pengucapan yang paling umum dari "satu". A
computer program attempting to decipher spoken Sebuah program komputer mencoba untuk
menguraikan diucapkan
English needs to provide for a variety of different ways that a word might be pronounced,
whereas a Bahasa Inggris perlu menyediakan berbagai cara yang berbeda bahwa sebuah kata
mungkin diucapkan, sedangkan
spelling checker needs only to know one. pemeriksa ejaan hanya perlu tahu satu.
An HMM is a variation on the well-known Markov chain model, one of the most widely Sebuah
HMM merupakan variasi dari rantai Markov dikenal model baik, salah satu yang paling banyak
studied stochastic models of discrete events (Bartholomew 1975). model stokastik mempelajari
kejadian diskrit (Bartholomew 1975). The standard reference on Referensi standar pada
HMMs is Rabiner (1989), which contains a thorough discussion of the estimation techniques
used HMMs adalah Rabiner (1989), yang berisi diskusi yang cermat tentang teknik estimasi yang
digunakan
with the models as well as setting forth a standard notation that is used in virtually all dengan
model serta yang mengatur suatu notasi standar yang digunakan dalam hampir semua
contemporary articles on the subject. artikel kontemporer pada subjek. Like a conventional
Markov chain, a HMM consists of a set Seperti rantai Markov konvensional, HMM terdiri dari
menetapkan
of n discrete states and an nxn matrix [A] = {a n keadaan diskrit dan matriks nxn [A] = {a
ij ij
} of      for going between }     terjadi antara
those states. negara-negara bagian. In addition, however, every state has a vector of   
    that Selain itu, bagaimanapun, setiap negara memiliki vektor  
   bahwa
combine into a second matrix [B] = { È bergabung menjadi a] kedua matriks [B = È
jj
(k)} corresponding to the probability that the system will (K)} sesuai dengan kemungkinan
bahwa sistem akan
produce a symbol of type k when it is in state j. menghasilkan simbol k tipe bila dalam negara j.
The states of the HMM cannot be directly Negara-negara dari HMM tidak bisa langsung
observed and can only be inferred from the observed symbols, hence the adjective "hidden".
diamati dan hanya dapat disimpulkan dari simbol-simbol yang diamati, maka kata sifat
"tersembunyi". This Ini
is in contrast to most applications of Markov models in international politics where the states
berbeda dengan sebagian besar aplikasi model Markov dalam politik internasional dimana
negara-negara
correspond directly to observable behaviors (see Schrodt 1985 for a review) berhubungan
langsung dengan perilaku diamati (lihat Schrodt 1985 untuk tinjauan)
While HMMs can have any type of transition matrix, the model that I will focus on in this
Sementara HMMs dapat memiliki semua jenis matriks transisi, model bahwa saya akan fokus di
dalam ini
chapter is called a "left-right model" because it imposes the constraint that the system can only
bab disebut "model kanan kiri" karena memaksakan kendala bahwa sistem hanya dapat
remain in its current state or move to the next state. tetap pada saat ini atau pindah ke negara
berikutnya. The transition matrix is therefore of the form Matriks transisi karena itu dalam
bentuk

a sebuah
11 11
1-a 1-a
11 11
0 ... 0 ...
00
00
a sebuah
22 22
1-a 1-a
22 22
... ...
00
00
00
a sebuah
33 33
... ...
00
... ...
... ...
00
00
0 ... 0 ... 1-a 1-a
n-1,n-1 n-1, n-1
00
00
0 ... 0 ...
11

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 4 Page 4
and the individual elements of the model look like those in Figure 1. dan elemen individu dari
model seperti yang terlihat pada Gambar 1. This model is widely used in Model ini banyak
digunakan dalam
speech recognition because parts of a word may be spoken slowly or quickly but in normal
speech pengenalan suara karena bagian dari kata mungkin diucapkan lambat atau cepat tetapi
dalam sambutannya normal
the ordering of those parts is never modified. urutan bagian-bagian tidak pernah diubah.
p  
È   
        
 È  
 
      
È  
  
 
a sebuah
i,i i, i
i,i+1 i, i +1
a sebuah
b (1) b (1)
b (m) b (m)
ii
ii
b (0) b (0)
ii

 
i,i-1 i, i-1
a sebuah
Recurrence Kambuh
probability kemungkinan
Transition Transisi
probabilities probabilitas
Symbol Simbol
probability kemungkinan
Observed Teramati
symbol simbol
   

A series of these individual elements form an HMM such as the 5-state model illustrated in
Figure Serangkaian unsur-unsur individu merupakan HMM seperti negara model 5 diilustrasikan
dalam Gambar
2. 2. This illustrates a ³left-right-left´ model, where a process can make a transition to the
previous Hal ini menggambarkan "kiri-kanan-kiri" model, dimana suatu proses dapat melakukan
transisi ke sebelumnya
state, the next state, or remain in the same state. negara, negara bagian berikutnya, atau tetap
dalam keadaan yang sama.
11
In empirical applications, the transition matrix and symbol probabilities of an HMM are Dalam
aplikasi empiris, transisi dan simbol matriks probabilitas sebuah HMM adalah
estimated using an iterative technique called the Baum-Welch algorithm. diestimasi dengan
teknik iteratif disebut algoritma Baum-Welch. This procedure takes a Prosedur ini membutuhkan
set of observed sequences (for example the word "seven" as pronounced by twenty different set
urutan diamati (misalnya kata "tujuh" sebagaimana diucapkan oleh dua puluh berbeda
speakers, or a set of dyadic interactions from the BCOW crisis set) and finds coefficients for the
speaker, atau satu set interaksi diad dari krisis BCOW set) dan menemukan koefisien untuk
matrices [A] and [B] that locally maximize the probability of observing those sequences. matriks
[A] dan [B] yang secara lokal memaksimalkan probabilitas mengamati urutan tersebut. The The
Baum-Welch algorithm is a nonlinear numerical technique and Rabiner (1989:265) notes "the
Baum-Welch algoritma adalah teknik numerik nonlinier dan Rabiner (1989:265) mencatat "yang
algorithm leads to a local maxima only and, in most problems of interest, the optimization
surface algoritma mengarah ke maxima lokal saja dan, dalam sebagian besar masalah
kepentingan, permukaan optimasi
is very complex and has many local maxima." sangat kompleks dan memiliki banyak maxima
lokal. "
11
This is a generalization of the ³left-right´ model commonly used in speech recognition, where
transitions are only Ini adalah generalisasi dari kiri-kanan "model" yang biasa digunakan dalam
pidato pengakuan, di mana transisi hanya
allowed to the next state. diizinkan untuk keadaan berikutnya. In a left-right model, the final state
of the chain is an "absorbing state" that has no Dalam model kanan kiri, rantai keadaan akhir
adalah sebuah "negara menyerap" yang tidak memiliki
exit probability and recurs with a probability of 1. keluar probabilitas dan berulang dengan
probabilitas 1.

c 
c 

Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan


Page 5 Page 5
p  
È          È      
   
È 

!!
""
##
$$
pp
Because the Baum-Welch algorithm is an expectation-maximization method, it should, in Karena
algoritma Baum-Welch adalah metode maksimisasi-harapan, harus, dalam
theory, be possible to use the standard tools of maximum likelihood methods to compute teori,
dimungkinkan untuk menggunakan alat-alat standar metode maksimum likelihood untuk
menghitung
asymptotic estimates of the standard errors of the estimates of the parameters in the [A] and [B]
asymptotic perkiraan kesalahan standar estimasi parameter dalam [A] dan [B]
matrices, as well as comparing different models using likelihood ratios. matriks, serta
membandingkan model yang berbeda menggunakan rasio kemungkinan. In practice, however,
this Dalam prakteknya, bagaimanapun, ini
does not seem to be done, at least in the literature I've surveyed. tampaknya tidak dilakukan,
setidaknya dalam literatur saya disurvei. The reason for this seems to be Alasan untuk ini
tampaknya
related to the local solutions provided by the Baum-Welch algorithm. terkait dengan solusi lokal
yang disediakan oleh algoritma Baum-Welch. As illustrated in the Seperti digambarkan dalam
experiments below, the variance of the parameter estimates found in these local solutions is very
eksperimen di bawah ini, varians parameter perkiraan ditemukan dalam solusi lokal sangat
large, although a variety of differing parameters appear to yield roughly similar estimates for the
besar, meskipun berbagai parameter yang berbeda muncul untuk menghasilkan perkiraan kasar
yang sama untuk
joint probability of the sequences. gabungan probabilitas urutan.
After a set of models has been estimated, that set can be used to classify an unknown sequence
Setelah set model telah diperkirakan, yang menetapkan dapat digunakan untuk
mengklasifikasikan sebuah urutan yang tidak diketahui
by computing the maximum probability that each of the models generated the observed
sequence. dengan menghitung probabilitas maksimum yang masing-masing model yang
dihasilkan urutan diamati.
This is done using an algorithm that requires on the order of N Hal ini dilakukan dengan
menggunakan suatu algoritma yang memerlukan atas perintah N
22
T calculations, where N is the T perhitungan, dimana N merupakan
number of states in the model and T is the length of the sequence. sejumlah negara dalam model
dan T adalah panjang urutan. Once the probability of the Setelah probabilitas dari
sequence matching each of the models is known, the model with the highest probability is chosen
as pencocokan urutan masing-masing model diketahui, model dengan probabilitas tertinggi
dipilih sebagai
that which best represents the sequence. yang paling mewakili urutan. Matching a sequence of
symbols such as those found in Yang cocok dengan urutan simbol seperti yang ditemukan di
daily data on a six-month crisis coded with using the 22-category World Events Interaction
Survey harian data pada bulan krisis enam dikodekan dengan menggunakan 22-kategori Acara
Dunia Interaksi Survei
scheme (WEIS; McClelland 1976), generates probabilities on the order of 10 skema (Weis;
McClelland 1976), menghasilkan probabilitas pada urutan 10
-(T+1) - (T +1)
: Assume that : Asumsikan bahwa
each state has ten associated WEIS categories that are equally probable: b setiap negara memiliki
sepuluh terkait Weis kategori yang sama-sama kemungkinan: b
ii
(k)=0.10. (K) = 0.10. Leaving aside Terlepas
the transition probabilities, each additional symbol will reduce the probability of the complete
probabilitas transisi, setiap simbol tambahan akan mengurangi probabilitas lengkap
sequence by a factor of 10 urutan dengan faktor 10
-1 -1
. . The transition probabilities, and the fact that the WEIS codes are not Probabilitas transisi, dan
fakta bahwa kode Weis tidak
equiprobable, further reduce this probability. equiprobable, mengurangi kemungkinan ini. These
sequence probabilities are consequently Ini probabilitas urutan secara konsekuen
   small, even if the sequence was in fact generated by one of the models, but the only
  kecil, bahkan jika urutan itu sebenarnya dihasilkan oleh salah satu model, tetapi hanya
important comparison is the   fit of the various models. perbandingan penting adalah fit
  dari berbagai model. The measure of fit usually reported is Ukuran fit biasanya dilaporkan
adalah
the log of the probability; this statistic is labeled i (alpha). log dari probabilitas, statistik ini diberi
label i (alpha).
(An insurmountable disadvantage of this computation is that one cannot meaningfully compare
(Sebuah kerugian dapat diatasi perhitungan ini adalah bahwa seseorang tidak dapat bermakna
dibandingkan
the fit of two sequences to a single HMM unless the sequences are equal in length. fit dari dua
sekuens ke HMM tunggal kecuali urutan yang sama panjang. In other words, Dengan kata lain,

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 6 Page 6
it is possible to compare a sequence to a series of models, but one cannot compare several
arbitrary adalah mungkin untuk membandingkan urutan ke serangkaian model, tetapi tidak dapat
membandingkan beberapa sewenang-wenang
sequences to a single model.) urutan ke model tunggal.)
For example, in a typical speech-recognition application such as the recognition of bank Sebagai
contoh, dalam aplikasi speech-recognition khas seperti pengakuan bank
account numbers, a system would have HMMs for the numerals "zero" through "nine". nomor
rekening, sistem akan HMMs untuk angka "nol" melalui "sembilan". When a Ketika
speaker pronounces a single digit, the system converts this into a set of discrete sound categories
pembicara mengucapkan digit tunggal, sistem mengkonversi ini menjadi satu set kategori suara
diskrit
(typically based on frequency), then computes the probability of that sequence being generated
by (Biasanya berdasarkan frekuensi), kemudian menghitung kemungkinan urutan yang yang
dihasilkan oleh
each of the ten HMMs corresponding to the ten digits. masing-masing sepuluh HMMs sesuai
dengan sepuluh digit. The HMM that has the highest The HMM yang tertinggi
probability²for example the HMM corresponding to the numeral "three"²gives the best
estimate probabilitas-misalnya HMM sesuai dengan angka "tiga"-memberikan perkiraan terbaik
of the number that was spoken. nomor yang diucapkan.
The application of the HMM to the problem of generalizing the characteristics of international
Penerapan HMM untuk masalah generalisasi karakteristik internasional
event sequences is straightforward. urutan peristiwa secara langsung. The symbol set consists of
the event codes taken from an Rangkaian simbol terdiri dari kode acara diambil dari
event data set such as WEIS. data peristiwa ditetapkan seperti Weis. The states of the model are
unobserved, but have a close theoretical Menyatakan model yang teramati, namun memiliki
dekat teoritis
analog in the concept of crisis "phase" that has been explicitly coded in data sets such as the
analog dalam konsep krisis "fase" yang telah secara eksplisit dikodekan dalam set data seperti
Butterworth international dispute resolution data set (Butterworth 1976), CASCON (Bloomfield
& penyelesaian sengketa internasional Butterworth data set (Butterworth 1976), CASCON
(Bloomfield &
Moulton 1989, 1997) and SHERFACS (Sherman & Neack 1993), and in work on preventive
Moulton 1989, 1997) dan SHERFACS (Sherman & Neack 1993), dan bekerja pada pencegahan
diplomacy such as Lund (1996). diplomasi seperti Lund (1996). For example, Lund (1996:38-39)
outlines a series of crisis Sebagai contoh, Lund (1996:38-39) menguraikan serangkaian krisis
phases ranging from "durable peace" to "war" and emphasizes the importance of an "unstable
fase mulai dari "perdamaian tahan lama" untuk "perang" dan menekankan pentingnya "tidak
stabil
peace" phase. In the HMM, these different phases would be distinguished by different
perdamaian "fase. Dalam HMM, fase-fase yang berbeda akan dibedakan oleh berbagai
distributions of observed WEIS events found in the estimated È distribusi peristiwa Weis diamati
ditemukan di È diperkirakan

vectors. vektor. A "stable peace" would Sebuah "perdamaian yang stabil" akan
have a preponderance of cooperative events in the WEIS  range; the escalation phase of the
memiliki dominan peristiwa koperasi di kisaran  Weis; fase eskalasi
crisis would be characterized by events in the % range (accusations, protests, denials, and
krisis akan ditandai dengan kejadian-kejadian di kisaran % (tuduhan, protes, penolakan, dan
threats), and a phase of active hostilities would show events in the & range. ancaman), dan
fase permusuhan aktif akan menunjukkan peristiwa di kisaran & The length of time Jangka
waktu
that a crisis spends in a particular phase would be proportional to the magnitude of the recurrence
bahwa krisis menghabiskan dalam tahap tertentu akan sebanding dengan besarnya kambuh
probability a probabilitas
ii ii
..
The HMM has several advantages over alternative models for sequence comparison. The HMM
memiliki beberapa keunggulan dibandingkan model-model alternatif untuk perbandingan urutan.
First, if Pertama, jika
N<<M, the structure of the model is relatively simple. N <<M, struktur model relatif sederhana.
For example a left-right model with N states Misalnya model kanan kiri dengan negara-negara N
and M symbols has 2(N-1) + N*M parameters compared to the M(M+2) parameters of a dan
simbol M memiliki 2 (N-1) + N * M parameter dibandingkan dengan parameter M (M +2) dari
Levenshtein metric. Levenshtein metrik. HMMs can be estimated very quickly, in contrast to
neural networks and HMMs dapat diperkirakan sangat cepat, berbeda dengan jaringan saraf dan
genetic algorithms. algoritma genetika. While the resulting matrices are only a local solution²
there is no guarantee Sedangkan matriks yang dihasilkan hanya lokal solusi-tidak ada jaminan
that a matrix computed by the Baum-Welch algorithm from a different random starting point
might bahwa matriks dihitung oleh algoritma Baum-Welch dari titik awal random yang berbeda
mungkin
be quite different²local maximization is also true of most other techniques for analyzing sangat-
lokal maksimisasi berbeda ini juga berlaku untuk sebagian besar teknik-teknik lain untuk
menganalisis
sequences. urutan. Furthermore, the computational efficiency of the Baum-Welch algorithm
allows Selain itu, efisiensi komputasi dari algoritma Baum-Welch memungkinkan
estimates to be made from a number of different starting points. estimasi harus dibuat dari
sejumlah titik tolak yang berbeda. The HMM model, being HMM model, yang
stochastic rather than deterministic, is specifically designed to deal with noisy input and with
stokastik bukan deterministik, dirancang khusus untuk menangani input berisik dan dengan
indeterminate time; both of these are present in international event sequences. waktu tak tentu;
kedua ini hadir dalam rangkaian acara internasional.

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 7 Page 7
HMMs are    ²model that characterizes a set of sequences can be HMMs
  m    yang mencirikan satu set urutan dapat
constructed without reference to the underlying rules used to code those sequences. dibangun
tanpa mengacu pada peraturan yang mendasari digunakan untuk kode tersebut urutan. This
provides Ini memberikan
a close parallel to the method by which human analysts generalize sequences: They typically
learn paralel dekat dengan metode dimana analis manusia generalisasi urutan: Mereka biasanya
belajar
general characteristics from a set of archetypal cases. umum karakteristik dari satu set kasus
tipikal.
HMMs do not require the use of interval-level scales such as those proposed by Azar and HMMs
tidak memerlukan penggunaan tingkat skala interval seperti yang diusulkan oleh Azar dan
Sloan (1975) or Goldstein (1992). Sloan (1975) atau Goldstein (1992). These scales, while of
considerable utility, assign weights to Skala ini, sementara dari utilitas yang cukup, menetapkan
bobot untuk
individual events in isolation and make no distinction, for example, between an accusation that
peristiwa individu dalam isolasi dan membuat tidak ada perbedaan, misalnya, antara tuduhan
yang
follows a violent event and an accusation during a meeting. mengikuti peristiwa kekerasan dan
tuduhan selama rapat. The HMM, in contrast, uses only the HMM, sebaliknya, hanya
menggunakan
original, disaggregated events and models the context of events by using different symbol asli,
terpilah peristiwa dan model konteks peristiwa dengan menggunakan simbol berbeda
observation probabilities in different states. observasi probabilitas di negara-negara yang
berbeda. An event that has a low probability within a particular Sebuah peristiwa yang memiliki
probabilitas rendah dalam waktu tertentu
context (that is, a specific hidden state) lowers the overall probability of the model generating the
konteks (yaitu, keadaan tersembunyi tertentu) menurunkan probabilitas keseluruhan model
menghasilkan
sequence. urutan. In aggregative scaling methods, events have the same weight in all contexts.
Dalam metode scaling agregatif, peristiwa memiliki berat yang sama dalam semua konteks.
While most existing work with event data aggregates by months or even years, the HMM
Sementara sebagian besar yang ada bekerja dengan data agregat acara oleh bulan atau bahkan
bertahun-tahun, yang HMM
requires no temporal aggregation. tidak memerlukan agregasi temporal. This is particularly
important for early warning problems, where Hal ini sangat penting untuk masalah peringatan
dini, di mana
critical periods in the development of a crisis may occur over a week or even a day. periode kritis
dalam pengembangan krisis mungkin terjadi selama seminggu atau bahkan sehari. The HMM is
HMM adalah
relatively insensitive to the delineation of the start of a sequence. relatif tidak sensitif dengan
deliniasi awal berurutan. It is simple to prefix an HMM Hal ini sederhana untuk awalan sebuah
HMM
with an initial "background" state that reflects the distribution of events generated by a particular
dengan "latar belakang" keadaan awal yang mencerminkan distribusi yang dihasilkan oleh
peristiwa tertentu
source (eg Reuters/WEIS) when no crisis is occurring. sumber (misalnya Reuters / Weis) bila
tidak ada krisis terjadi. A model can simply cycle in this state Sebuah model dapat hanya siklus
dalam keadaan ini
until something important happens and the chain moves into the later states characteristic of
crisis sampai sesuatu yang penting terjadi dan bergerak rantai ke negara kemudian karakteristik
krisis
behavior. perilaku.
There is a clear interpretation to each of the parameters of the [A] and [B] matrices, which Ada
interpretasi yang jelas untuk setiap parameter [A] dan [B] matriks, yang
allows them to be interpreted substantively; this contrasts with techniques such as neural
networks memungkinkan mereka untuk ditafsirkan secara substansial, ini kontras dengan teknik
seperti jaringan saraf
that have a very diffuse parameter structure. yang memiliki parameter struktur difus sangat.
More generally, the fit of the model has a familiar Lebih umum, kesesuaian model memiliki
akrab
probabilistic interpretation. interpretasi probabilistik. Finally²and not insignificantly²the
HMM technique has already Akhirnya-dan bukan tidak signifikan-teknik HMM telah
been developed and is an active research topic in a number of different fields. telah
dikembangkan dan merupakan topik penelitian yang aktif dalam berbagai bidang yang berbeda.
The breadth of those Luasnya tersebut
applications indicates that the method is relatively robust. aplikasi menunjukkan bahwa metode
ini relatif kuat. While there is always a danger in Sementara selalu ada bahaya di
applying the  m     to whatever data on political behavior happen to be laying
around, menerapkan      data apapun pada perilaku politik yang terjadi untuk menjadi
peletakan sekitar,
the HMM appears unusually well suited to the problems of generalizing and classifying HMM
muncul luar biasa cocok untuk masalah generalisasi dan mengklasifikasikan
international event data sequences. data event internasional urutan.
Data and Forecasting Model Data dan Model Peramalan
Data Data
The event data used in this study were machine-coded using the 2-digit (22 category) WEIS data
peristiwa yang digunakan dalam penelitian ini adalah mesin-kode menggunakan 2 digit (22
kategori) Weis
system from the lead sentences in Reuters stories obtained from the NEXIS data service for the
sistem dari kalimat utama dalam cerita Reuters diperoleh dari layanan data Nexis untuk
period January 1991 through May 1997 and the Reuters Business Briefing service for June 1997
periode Januari 1991 sampai Mei 1997 dan Reuters layanan Business Briefing untuk Juni 1997

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 8 Page 8
through January 1999. hingga Januari 1999. These reports were coded using the Kansas Event
Data System (KEDS) Laporan ini diberi kode dengan menggunakan Kansas Event Data System
(Keds)
automated event data coding program (Gerner et al. 1994; Schrodt, Davis & Weddle 1994).
otomatis acara pengkodean data program (Gerner et al;. 1994 Schrodt, Davis & Weddle 1994).
The KEDS coder does some simple linguistic parsing of the news reports²for instance, it Koder
Keds melakukan beberapa parsing linguistik sederhana dari berita laporan-misalnya,
identifies the political actors, recognizes compound nouns and compound verb phrases, and
mengidentifikasi para aktor politik, mengakui nomina majemuk dan frasa verba majemuk, dan
determines the references of pronouns²and then employs a large set of verb patterns to
determine menentukan referensi dari kata ganti-dan lalu menggunakan satu set besar pola verba
untuk menentukan
the appropriate event code. acara kode yang sesuai. Only the lead sentences were coded and a
sentence was not coded if it Hanya kalimat memimpin diberi kode dan kalimat tidak kode jika
contained six or more verbs or no actor was found prior to the verb (sentences meeting these
berisi enam atau lebih verba atau aktor tidak ditemukan sebelum kata kerja (pertemuan kalimat
ini
criteria have a greater-than-average likelihood of being incorrectly coded by KEDS). kriteria
memiliki dari rata-rata kemungkinan besar menjadi salah kode oleh Keds). Schrodt & Schrodt &
Gerner (1994), Huxtable & Pevehouse (1996) and Bond et al. Gerner (1994), Huxtable &
Pevehouse (1996) dan Obligasi et al. (1996) discuss extensively the (1996) membahas secara
ekstensif
reliability and validity of event data generated using Reuters and KEDS. reliabilitas dan validitas
data peristiwa dihasilkan menggunakan Reuters dan Keds. A  nonevent was added Sebuah
nonevent  ditambahkan
for each day in which no events were recorded in either direction in the dyad. untuk setiap hari
dimana ada kegiatan dicatat di kedua arah di angka dua itu. Multiple events Beberapa peristiwa
occurring in the same day are kept in the sequence. terjadi pada hari yang sama disimpan dalam
urutan. While the JWAC coding dictionaries sub- Sementara JWAC pengkodean sub-kamus
divide the various geographical-ethnic groups in the Balkans²for example separately coding
membagi kelompok geografis-berbagai etnis di Balkan-misalnya secara terpisah coding
Bosnian Serbs²these actors were combined into four primary ethnic groups²Serbs, Croats,
Serbia-Bosnia aktor ini digabung menjadi empat kelompok etnis utama-Serbia, Kroasia,
Bosnians, and Kosovars²for the purpose of the analysis. Bosnia, dan Kosovo-untuk tujuan
analisis.
Because Reuters was intensely covering this region during most of the period analyzed, only
Karena itu sangat Reuters meliput daerah ini selama sebagian besar dari periode yang dianalisis,
hanya
lead sentences were coded. kalimat memimpin diberi kode. The KEDS program is capable of
coding full stories and has been Program Keds yang mampu coding cerita penuh dan telah
tested in that capacity (see Schrodt and Gerner 1998), but we have found that full-story coding
diuji dalam kapasitas (lihat Schrodt dan Gerner 1998), tapi kami telah menemukan bahwa penuh
cerita coding
gives additional information only when marginal actors are involved, or when an area receives
only memberikan informasi tambahan hanya jika aktor marjinal yang terlibat, atau ketika suatu
daerah hanya menerima
sporatic coverage (for example West Africa; see Huxtable 1997). sporatic cakupan (misalnya
Afrika Barat, lihat Huxtable 1997). Neither condition applies to the Kondisi Baik berlaku untuk
Balkans: significant events in the conflict almost inevitably received coverage in separate stories
Balkan: peristiwa penting dalam konflik pasti mendapat liputan hampir dalam cerita-cerita
terpisah
(ie in lead sentences), and in many cases, a single event would generate multiple stories. (Yaitu
dalam kalimat memimpin), dan dalam banyak kasus, sebuah peristiwa tunggal akan
menghasilkan beberapa cerita.
Full-story coding would primarily serve to insert a large number of high-frequency verbal
Kendali-cerita coding terutama akan berfungsi untuk memasukkan sejumlah besar frekuensi
tinggi verbal
events ("comment" and "consult") into the sequences, and because the HMM models use
peristiwa ("komentar" dan "berkonsultasi") ke dalam urutan, dan karena model HMM
menggunakan
templates that contain a fixed number of events (this is necessitated by the HMM approach), the
template yang tetap mengandung sejumlah peristiwa (ini adalah diharuskan oleh pendekatan
HMM), yang
effect of full-story coding would be to  m the amount of information in any given sequence.
efek penuh cerita akan coding untuk     jumlah informasi dalam urutan yang diberikan.
22
By using lead-sentence coding, we insure that the sequences contain the most important events
Dengan menggunakan timbal-kalimat coding, kami memastikan bahwa urutan berisi peristiwa
yang paling penting
(subject, as always, to the judgements of Reuters' reporters and editors.) (Subjek, seperti biasa,
untuk penilaian 'wartawan Reuters dan editor.)
22
That is, reduce the amount of information from the perspective of information theory: because
the "comment" Artinya, mengurangi jumlah informasi dari perspektif teori informasi: karena
"komentar"
category in WEIS is very frequent, it carries less information than an event that occurs with low
probability. kategori Weis sangat sering, ia membawa informasi kurang dari suatu peristiwa yang
terjadi dengan probabilitas rendah.

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 9 Page 9
Forecasting Model Model Peramalan
The accuracy of the model is accessed on whether it can predict when violence will occur,
Ketepatan model diakses pada apakah ia dapat memprediksi kapan kekerasan akan terjadi,
violence being defined as events coded into the WEIS "22" category. kekerasan yang
didefinisikan sebagai peristiwa kode ke dalam "22" kategori Weis. Given the diverse character
Mengingat karakter beragam
of conflict in this region over the time period²violence having occurred between Serbs and
Croats, konflik di wilayah ini selama periode waktu-kekerasan yang terjadi antara Serbia dan
Kroasia,
Serbs and Bosnians, Serbs and Kosovars, but also occasionally between other groups (eg Croats
Serbia dan Bosnia, Serbia dan Kosovo, tetapi juga kadang-kadang antara kelompok-kelompok
lain (misalnya Kroasia
and Bosnians)² I did not attempt to differentiate who was involved in the violence. dan Bosnia)
- Saya tidak berusaha untuk membedakan yang terlibat dalam kekerasan.
The forecast target was whether a week contained more or less than twenty (20) WEIS Target
perkiraan adalah apakah seminggu berisi lebih atau kurang dari dua puluh (20) Weis
category  ("use of force") events. kategori  ("penggunaan kekuatan") peristiwa. The
threshold of 20 events is somewhat arbitrary, but that Ambang 20 peristiwa agak sewenang-
wenang, tetapi
threshold seemed to be reasonable for differentiating periods when the Balkans were relatively
ambang tampaknya masuk akal untuk membedakan periode ketika Balkan relatif
quiet from those where there was substantial violence. tenang dari orang-orang di mana ada
kekerasan substansial. Approximately 20% of the weeks in the Sekitar 20% dari minggu dalam
data set satisfy the "high conflict" criterion. set data memenuhi "konflik tinggi" kriteria.
p ' 
È ' ( )  )
 (  

Time Waktu
Start of Mulai
target target
week minggu
7 days 7 hari
Forecasting lag Peramalan lag
(28, 91 or 184 days) (28, 91 atau 184 hari)
100-event 100-acara
predictive prediksi
sequence urutan
Measurement Pengukuran
of WEIS 22 of Weis 22
events peristiwa
End of Akhir
sequence urutan
Three forecast periods were used Tiga periode proyeksi yang digunakan
28 days 28 hari
approximately 1 month sekitar 1 bulan
91 days 91 hari
approximately 3 months sekitar 3 bulan
184 days 184 hari
approximately 6 months sekitar 6 bulan
The early warning sequence for each week consisted of the 100 events prior to the first day of
Urutan peringatan dini untuk setiap minggu terdiri dari 100 peristiwa sebelum hari pertama
the week minus the forecast period; minggu minus periode peramalan;
33
Figure 3 provides a schematic of this model. Gambar 3 memberikan skematik model ini.
33
Why 100?²because I have ten fingers... Kenapa 100?-Karena saya punya sepuluh jari ... The
length of the warning sequence is a free parameter²SEQ_SIZE in Panjang urutan peringatan
adalah SEQ_SIZE bebas-parameter dalam
the program²and other values might work better, depending on the application. program-dan
nilai-nilai lain mungkin bekerja lebih baik, tergantung pada aplikasi. In previous work on the
Dalam pekerjaan sebelumnya di
Middle East, I have done some experiments with sequences of 50 and 200 events; the results
were roughly Timur Tengah, saya telah melakukan beberapa eksperimen dengan urutan dan 200
kejadian 50; hasilnya kasar
comparable to the results from 100 event sequences. sebanding dengan hasil dari 100 urutan
acara. Given the vagaries of timing in this region²for example Mengingat keanehan waktu di
wilayah ini-misalnya
the unpredictable and seasonally variable effects of weather on military operations, as well as the
large number of yang tidak terduga dan musiman variabel pengaruh cuaca di operasi militer,
serta sejumlah besar

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 10 Page 10
The HMMs were estimated using 16 "forecasting templates," eight for the high-conflict weeks
Para HMMs diperkirakan dengan menggunakan 16 "peramalan template," delapan untuk konflik
minggu tinggi
and eight for the low-conflict weeks. dan delapan untuk konflik minggu rendah. A high-conflict
template was created from the data by A-konflik template tinggi diciptakan dari data dengan
choosing a high-conflict week at random, then getting a 100-event sequence with the appropriate
memilih-konflik minggu tinggi secara acak, kemudian mendapatkan-acara urutan 100 dengan
yang sesuai
forecasting lag; low-conflict templates are created in a similar manner. peramalan lag; rendah
konflik template dibuat dengan cara yang sama. Finally, prediction for a Akhirnya, prediksi
untuk
particular week is done by computing the fit of a 100-event sequence, obtained with a suitable
lag minggu tertentu dilakukan dengan menghitung sesuai dari urutan 100-event, yang diperoleh
dengan lag yang sesuai
prior to the beginning of the week, and then assigning the "high conflict" or "low conflict"
prediction sebelum awal minggu, dan kemudian menempatkan "konflik tinggi" atau "konflik
rendah" prediksi
depending on which one of the two models had a higher probability of generating the sequence.
tergantung pada salah satu dari dua model memiliki kemungkinan lebih tinggi untuk
menghasilkan urutan.
In this preliminary analysis, the templates have been chosen at random from   week in the
Dalam analisis awal, template telah dipilih secara acak dari minggu   di
data set, so while the computed   m is strictly predictive (ie the prediction technique uses no
set data, jadi sementara   dihitung adalah sangat prediktif (yaitu tidak menggunakan
teknik prediksi
information beyond that available at the beginning of the week minus FORECAST days), the
informasi di luar yang tersedia pada awal minggu minus hari PRAKIRAAN), yang
 of the predictive model is not, because the model can use templates that occurred after
 model prediksi tidak, karena model dapat menggunakan template yang terjadi setelah
the time of the forecast. waktu ramalan. (This approach was used because I anticipated that some
periods of time (Pendekatan ini digunakan karena saya mengantisipasi bahwa beberapa periode
waktu
would provide higher-quality templates than others, though this turned out not to be the case so
far akan menyediakan template berkualitas tinggi daripada yang lain, meskipun ini ternyata tidak
menjadi kasus sejauh
as I can determine). seperti yang saya dapat menentukan). Nonetheless, only a very small amount
of the data Meskipun demikian, hanya jumlah yang sangat kecil data
44
²approximately 4%²is -Sekitar 4%-adalah
being used to characterize the complete sequences, so the model is certainly not tautological.
yang digunakan untuk mencirikan urutan yang lengkap, sehingga model ini tentu tidak
tautologis. It is Hal ini
straightforward to switch the estimation is a purely predictive mode; this will be done for the
final mudah untuk beralih estimasi adalah murni prediktif mode, ini akan dilakukan untuk final
analysis in this project once the various elements of the model have been finalized. analisis
dalam proyek ini setelah berbagai elemen model telah dituntaskan. Two additional Dua
tambahan
estimation schemes that are purely predictive²they estimate the model on a set of data that
occurs estimasi skema yang murni prediktif-mereka mengestimasi model pada satu set data yang
terjadi
before the period where the accuracy is assessed²are also estimated; they will be discussed
sebelum masa dimana akurasi dinilai-juga diperkirakan, mereka akan dibahas
below. di bawah ini.
The forecasting model used the following eight relationships Model peramalan yang digunakan
hubungan berikut delapan
Serbia -> any target Serbia -> sasaran apapun
Croatia -> any target Kroasia -> sasaran apapun
any source -> Serbia sumber -> Serbia
any source -> Croatia sumber -> Kroasia
Bosnia -> any target Bosnia -> sasaran apapun
Kosovo -> any target Kosovo -> sasaran apapun
any source -> Bosnia sumber -> Bosnia
any source -> Kosovo sumber -> Kosovo
Following the approach in Schrodt (2000), the multiple interactions were modeled by
incrementing Mengikuti pendekatan dalam Schrodt (2000), beberapa interaksi dimodelkan oleh
incrementing
the WEIS code for the Nth dyad by (N-1)*22, so for example the {any source -> Serbia} events
kode Weis untuk Nth angka dua oleh (N-1) * 22, jadi misalnya {sumber -> Serbia} kejadian
have codes ' through ** (corresponding to the original WEIS codes  to  ) the {Croatia ->
memiliki kode ' melalui ** (sesuai dengan kode Weis asli  fungsi {Kroasia ->
any target} events have codes &+ through  , and so forth. setiap target} kejadian memiliki
kode &+ melalui , dan sebagainya. If no event occurred with either dyad, Jika acara tidak
terjadi dengan angka dua baik,
diplomatic interventions during the course of the conflict²it is unlikely that the model will be
very sensitive to intervensi diplomatik selama konflik-itu tidak mungkin bahwa model akan
sangat sensitif terhadap
the length of the sequence. panjang urutan.
44
1600 events in the templates out of about 42,000 events in the total sequence. 1600 peristiwa
dalam template dari sekitar 42.000 kejadian dalam urutan total.

c  c 
Schrodt: Forecasting Conflict Schrodt: Konflik Peramalan
Page 11 Page 11
the  nonevent was assigned to the day. yang nonevent  ditugaskan untuk hari ini. The
resulting model contains 177 event codes (8*22 + Model yang dihasilkan berisi kode acara 177
(8 * 22 +
1). 1).
This approach involves a relatively high level of aggregation, since it does not differentiate the
Pendekatan ini melibatkan tingkat tinggi yang relatif agregasi, karena tidak membedakan
source or target of interactions with the four principals in the conflict. sumber atau target
interaksi dengan empat pelaku dalam konflik tersebut. This was done for three Hal ini dilakukan
selama tiga
reasons. alasan. First, if I had tried to further differentiate sources and targets, the number of
codes Pertama, jika saya harus mencoba untuk lebih membedakan sumber dan target, jumlah
kode
required in the model would have increased geometrically: For example looking only at the four
diperlukan dalam model akan meningkat secara geometris: Misalnya melihat hanya pada empat
principals plus the international community would have involved 20 dyads, which would mean
441 ditambah kepala masyarakat internasional akan melibatkan 20 diad, yang berarti 441
codes, which leads to a parameter space of dimension 5,324, which makes consistent estimation
kode, yang mengarah ke ruang parameter dimensi 5324, yang membuat estimasi konsisten
that much more difficult. yang jauh lebih sulit. Second, because of th Serbia first Croatia, then
Bosnia, then a Croat- Kedua, karena pertama Serbia Kroasia th, kemudian Bosnia, maka
Kroasia-
Bosnian alliance,a template chosen at any particular time would primarily consist of zero values
for aliansi Bosnia, template yang dipilih pada waktu tertentu terutama akan terdiri dari nilai nol
untuk
almost all of the dyads, and therefore would not generalize easily. hampir semua diad, dan
karenanya tidak akan menggeneralisasi mudah. This approach, in contrast, Pendekatan ini,
sebaliknya,
essentially asks "What do Serbian (or Croatian, Bosnia, Kosovar) patterns look like when Serbia
is dasarnya bertanya "Apa Serbia (atau Kroasia, Bosnia, Kosovo) pola terlihat seperti ketika
Serbia
interacting with someone, anyone." Finally, the specific sources and targets are generally
consistent berinteraksi dengan seseorang, siapa pun Akhirnya. ", sumber spesifik dan target
umumnya konsisten
during a given pef time, so while the dyadic behaviors are aggregated, it is simple to identify the
selama waktu PEF diberikan, sehingga sementara perilaku diadik dikumpulkan, adalah sederhana
untuk mengidentifikasi
actors responsible for most of the activity at any given point in time. waktu aktor bertanggung
jawab atas sebagian besar aktivitas pada diberikan setiap titik.
Estimation Algorithm Estimasi Algoritma
The HMM parameters were estimated by extensively modifying the source code written by
Parameter HMM diestimasi oleh luas memodifikasi kode sumber yang ditulis oleh
Meyers & Whitson (1995). Meyers & Whitson (1995). Their C++ code implements a left-right
hidden Markov model and Mereka C + + kode mengimplementasikan hak-model Markov
tersembunyi kiri dan
the corresponding Baum-Welch maximum likelihood training algorithm. maksimum sesuai
Baum-Welch pelatihan kemungkinan algoritma. I translated this code Aku menerjemahkan kode
ini
from the Solaris C++ environment to a Macintosh CodeWarrior ANSI C environment, in the dari
+ + C lingkungan Solaris ke CodeWarrior ANSI C lingkungan Macintosh, di
process combining Meyers and Whitson's separate driver programs for training and testing into a
menggabungkan proses Meyers dan terpisah pengemudi program Whitson untuk pelatihan dan
pengujian ke
single program, and modifying the input format to handle the WEIS sequences. program tunggal,
dan memodifikasi format masukan untuk menangani urutan Weis. I then extended Saya
kemudian diperpanjang
the code to handle the left-right-left (LRL) model, and implemented the Viterbi algorithm
described kode untuk menangani kiri (LRL) model-kanan-kiri, dan menerapkan algoritma
Viterbi dijelaskan
in Rabiner (1989) in order to estimate the most likely state sequence. di Rabiner (1989) untuk
memperkirakan urutan negara yang paling mungkin. In the process of extending Dalam proses
perpanjangan
the model to the LRL form, I rewrote the estimation equations to correspond exactly to those in
Rabiner²the Meyers & Whitson implementation differed slightly from Rabiner's equations,
presumably because their models estimate a separate vector for "transition symbols." These new
Ini baru
procedures produce estimates similar to those of Meyers & Whitson when all probabilities to
previous states are forced to zero. This source code is available on the KEDS project web site:
http://www.ukans.edu/keds/software.html.
The program requires about 2 Mb of memory for a system using 177 codes, 6 states and 100-
event sequences. The largest arrays required by the program are proportional to (M+T)*N, where
M is the number of possible event codes, T is the maximum sequence length and N is the number

c  c 
Schrodt: Forecasting Conflict
Page 12 Page 12
of states. negara. Consistent with the CASCON and SHERFACS approaches, and with earlier
work that I
did on the Middle East (Schrodt 1999, 2000), the models I estimated used 6 states.
On average, estimating a single 177-code, 6-state HMM model using 8 template sequences takes
about 0.80 seconds on a 350 Mhz Apple Macintosh G3, and calculating the probability of a given
sequence²ie doing a prediction²is practically instantaneous. Unfortunately, this speed is
canceled out by the fact that the Baum-Welch algorithm is not very efficient at finding global
maxima, and therefore a combination of Monte Carlo and genetic algorithm methods were
required
to locate good models [see discussion below]. In the protocols I've been using, a standard run
requires estimating 49,664 models (= 2* {[(24*64)+16]*16}) so it takes around 11 hours. This
Ini
limited full evaluation to a couple of runs per day on a personal, so it was difficult to experiment
with
a large number of parameters. However, I would reiterate that this is only a constraint on
 
the optimal model²once models have been established, new data can be evaluated against these
very,
very quickly. sangat cepat.
[In the course of doing this research, I adapted the programs to run on a parallel supercomputer
at the National Center for Supercomputer Applications; this experience is described at
http://www.ukans.edu/keds/NCSA.html . The supercomputer definitely allowed greater
experimentation; these results are briefly summarized below.]
As noted above, the optimization surface of an HMM calculated using the Baum-Welch
algorithm is characterized by a very large number of local maxima. This means that the resulting
the parameters²and the predictive accuracy²of the model depend heavily on the initial
approximation for those parameters. In other words, the point where the estimation routine stops
depends heavily on where it starts, and those starting points occur in a 2156-dimensional space
²2 models * [(6 states * 177 codes) + 16 transition probabilities].
In order to efficiently explore this surface, I used a combination of Monte Carlo methods and
genetic algorithms. algoritma genetika. This was done as follows (the capitalized variable names
correspond to
constants in the estimation program):
1. 1. A set of GA_VECT initial starting points for a pair of models²one for the high conflict
templates and one for the low conflict templates²was chosen (typically GA_VECT = 32).
The initial parameter values were chosen from a uniform [0,1] distribution, then normalized
so that they summed to 1.
2. 2. An HMM was estimated for each of these starting points, then the "fitness" of the estimated
model was computed. "Fitness" was some combination of correct minus incorrect
predictions (see below).

c  c 
Schrodt: Forecasting Conflict
Page 13 Page 13
3. 3. After all of the models had been computed, the vectors were sorted in descending order of
fitness. kebugaran. Only the top GA_SURVIVE (=0.25) * GA_VECT vectors were saved.
4. 4. The remaining (1 - GA_SURVIVE ) * GA_VECT vectors were replaced by reproducing the
surviving vectors using the standard genetic algorithm method: two "parents" were chosen at
random (the likelihood of choosing being proportional to their fitness), then a new vector
was created by taking the first NR elements of the first vector (where NR is a random
number) and the remaining elements from the second vector. In addition, elements were
randomly mutated with probability GA_MUTATE (=0.10).
This process was repeated for GA_MAX_EON (=64) iterations, then the best-fitting model was
saved, and a set of predictions was made using that best-fitting model. Finally, the GA was
repeated with a number of different starting points and templates (GA_EXPER, which was set at
16 but repeated multiple times in some cases).
The GA seems to work fairly efficiently, though it does not converge to a single model in the
Balkans case. However, the range of optimal values is relatively limited, typically varying by
around 15% - 20% between the best and worst models. This variation, while not ideal, is
considerably more limited than what I got using a simpler Monte-Carlo estimation.
55
The GA is
not, however, a ³magic bullet´ and is itself quite time-consuming²I did a small number of
experiments where I estimated models using a pure Monte-Carlo search that involved the same
number of evaluations as the GA, and the GA gives only about a 10%-20% improvement in
accuracy for the same amount of work. Other complex optimization approaches such as
simulated
annealing might produce better results.
While I used GA_MAX_EON =64 in virtually all of the experiments, most of the optimization
tends to occur early in the process.
66
I did some experiments (on a simpler model, Israel-Lebanon)
with altering the various parameters of the GA²GA_VECT, GA_SURVIVE and
GA_MUTATE²and the optimization doesn't seem very sensitive to these values.
77
(I did a run of 16
55
This variation is partly due to the different choice of templates for high and low conflict periods:
some templates
are going to be more "typical" than others, and consequently will provide a better fit. In a
separate experiment
on Israel-Lebanon data where a fixed set of templates was used, the variation among experiments
was in the
range of 10%.
66
Virtually all of the optimization occurs in the first 32 eons of the genetic algorithm, and one
could probably set
GA_MAX_EON =32 with almost no loss of accuracy. This would increase the speed of the
experiments by a
factor of 2. faktor 2.
77
I did a run of 16 experiments on the 3N model for the Balkans data with GA_VECT=48 and
GA_SURVIVE=0.33²in other words, keeping 16 rather than 8 of the "parent" vectors²and this
made no
discernible difference in either the overall accuracy or the rate of convergence.

c  c 
Schrodt: Forecasting Conflict
Page 14 Page 14
experiments on the 3N model for the Balkans data with GA_VECT=48 and
GA_SURVIVE=0.33
²in other words, keeping 16 rather than 8 of the "parent" vectors²and this made no discernible
difference in either the overall accuracy or the rate of convergence.)
Alternative Measures of Fitness
The fitness of the model-pairs is computed using
™Ȉ
(w
TP TP
+w+W
TN TN
²w
FP FP
²w
FN FN
))
where the w
aa aa
correspond to the weights for a
TP: true positive
high conflict predicted when high conflict occurs
TN: true negative
low conflict predicted when low conflict occurs
FP: false positive
high conflict predicted when low conflict occurs
FN: false negative
low conflict predicted when high conflict occurs
This is equivalent to a weighted "right minus wrong" criterion:
(#correct predictions) minus (#incorrect predictions)
The sum is over all of the weeks in the data set (from 1 January 1991 to 26 January 1999); each
week is classified into one of these categories.
Because the data set is strongly skewed towards low-conflict weeks (80% of the cases), this
model will generally under-predict high-conflict: In particular a null model that predicts 
low-
conflict for all weeks would have an impressive 80% accuracy, but it will also be quite useless.
This Ini
is the perennial early warning problem of balancing Type I and Type II errors: how should a
model
balance the possibility of false alarms with the possibility of missing actual cases of high
conflict.
One can create a [useless] model with zero false alarms by   predicting high conflict, and
one
can also create an equally useless model that misses none of the cases of high conflict by 
predicting high conflict. There is no simple way around this tradeoff.

c 
c 

Schrodt: Forecasting Conflict


Page 15 Page 15
The simplest way of doing this is to differentiate reward and penalize some types of predictions
over others. atas orang lain. I experimented with three different weighting systems:
ww
TP TP
ww
TN TN
ww
FP FP
ww
FN FN
Unweighted
1.0 1.0
1.0 1.0
1.0 1.0
1.0 1.0
Entropy, FP
1.68 1.68
0.21 0.21
1.68 1.68
0.21 0.21
Entropy, FN
1.68 1.68
0.21 0.21
0.21 0.21
1.68 1.68
The entropy-based weights are suggested by the information theory computation of the entropy
of a
sequence urutan
E=-
™Ȉ
i=1 i = 1
CC
pp
ii
ln(p ln (p
ii
))
where mana
pp
ii
= the proportion of category i in the data and
C = the number of distinct categories.
and ln() is the natural logarithm. By taking only the log of the proportion, one gets the weights
-ln( 0.186) = 1.68
high conflict
-ln( 0.814) = 0.21
low conflict
These weights were used as the "reward" for a correct prediction for each type of week. I also
Saya juga
looked at two different ways of weighting the incorrect predictions: one using the high-conflict
weight to penalize false positives; the other uses the high-conflict weight on false negatives.
88
Results Hasil
Models were estimated for the three forecasting horizons and the three weighting systems, a
total of nine models. At least sixteen Monte-Carlo experiments were run for each model; in some
cases there are a larger number of models because I was able to salvage results from assorted
experiments. percobaan.
In all of the tables and figures, the numerical prefix (1, 3, and 6) refers to the forecasting
horizon; the letter prefix (U, P, N) refers to the fitness weighting. In some cases, I analyze
88
There is no particular reason that the penalty for incorrect predictions needs to have the same
value as the reward
for correct predictions, but this will do for a first approximation.

c  c 
Schrodt: Forecasting Conflict
Page 16 Page 16
separately the best-fitting models (accuracy ë 0.795 for U and P models; accuracy ë 0.495 for N
models) and for all models. Unless otherwise noted, the discussion will refer to the "best"
models,
but the "all" results are presented for purposes of comparison and to give some sense of the
variance found in the estimates. In all of the analyses, the U and P models produce very similar
results, with the N model being distinctive.
Overall Accuracy
The overall accuracy of the models is shown in Tables 1 and 2; additional tables show the four-
way classification accuracy. Table 1 summarizes the number of models evaluated, the total
number
of observations (ie total number of weeks predicted across all of the models), and the average
accuracy (true positive + true negatives). The key result here is that the overall accuracy for the
best U and P models consistently show almost exactly 80% accuracy; the best N models
somewhat less consistently show about 52% accuracy. The difference between the "best" and
"all"
model sets
99
is not dramatic²only about 5%²and quite surprisingly, there is very little drop-off in
accuracy as the time horizon increases.
[This last characteristic worries me a bit, but I have thoroughly checked through the program
code to make sure that this is not an artifact, and I can't find any errors. In addition, I have also
used these programs to estimate models for a number of other conflict regions and the pattern of
small decreases in accuracy²rather than accuracy increasing and decreasing randomly²is
found
consistently. konsisten. The reason for this lack of sensitivity to time lags may be due to the
episodic
character of violence in the Balkans and elsewhere. The data set is characterized by two extended
high-conflict periods²May-93 to June-94 and April-95 to October-95²with the remainder of
the
period having only sporadic conflict, often lasting only a couple of weeks. These gross
characteristics may be predictable quite far in advance.]
99
"All" includes the "Best" models

c  c 
Schrodt: Forecasting Conflict
Page 17 Page 17
ã È  ã È 


$
 
!   ãÈ 
  
 
 - 
-.È
/") /! 
- 
-.È
/")
1-U
10 10
4090
80.7% 80,7%
22 22
8998 8998
78.5% 78,5%
3-U
77
2800 2800
80.6% 80,6%
21 21
8400 8400
76.9%
6-U
22
772 772
81.2% 81,2%
16 16
6176 6176
75.9% 75,9%
1-P 1-P
77
2863 2863
80.8% 80,8%
29 29
11861 11861
77.6% 77,6%
3-P 3-P
11 11
4400 4400
80.8% 80,8%
54 54
21600 21600
76.0% 76,0%
6-P 6-P
33
1158 1158
80.0% 80,0%
16 16
6176 6176
76.9%
1-N
15 15
6135 6135
55.2% 55,2%
16 16
6544 6544
54.2% 54,2%
3-N
88
3200 3200
52.8%
16 16
6400 6400
49.0% 49,0%
6-N
77
2702 2702
53.7% 53,7%
16 16
6176 6176
47.7% 47,7%
Tables 2a and 2b show the accuracy of the forecasts broken down by high-conflict and low-
conflict weeks respectively. The "observed" column gives the percentage of the weeks that were
correctly forecast : this proportion is
TP TP
TP+FN
for high conflict and
TN TN
TN+FP
for low conflict. It is Hal ini
the percentage of time that a high or low conflict week would have been predicted correctly.
The "forecast" column, in contrast, gives the percentage of the weeks that were forecast as
having high or low conflict actually turned out to have the predicted characteristic. This is Ini
adalah
TP TP
TP+FP
for high conflict and
TN TN
TN+FN
for low conflict. It is the percentage of time that a type of prediction
is accurate.
As indicated in the discussion of the weighting systems, the N-type models operate very
differently than the U- and P-type models. As shown in Table 1, U and P models have a high
overall accuracy. However, this accuracy comes almost entirely from correctly forecasting low-
conflict weeks²U and P models predict about 95% of these weeks correctly, but correctly
predict
only about 25% of the high-conflict weeks. N-type models, in contrast, predict about 85% of the
high-conflict weeks correctly (and 45% of the low-conflict weeks).

c  c 
Schrodt: Forecasting Conflict
Page 18 Page 18
ã È  ã È 
0 "  )1
!   ãÈ 
  
 
 
.È ã

p)  

.È ã

p)  

1-U
24.2% 24.2%
55.5% 55,5%
31.3% 31,3%
45.7% 45,7%
3-U
26.9% 26,9%
57.1%
27.9% 27,9%
41.6% 41,6%
6-U
30.1% 30,1%
63.3% 63,3%
32.3% 32,3%
42.2% 42,2%
1-P 1-P
21.3% 21,3%
55.0% 55,0%
29.3% 29,3%
40.8% 40,8%
3-P 3-P
22.7% 22,7%
57.1%
29.0% 29,0%
37.9% 37,9%
6-P 6-P
24.3% 24,3%
57.1%
25.9% 25,9%
42.6% 42,6%
1-N
92.8%
28.6% 28,6%
92.67% 92,67%
28.1% 28,1%
3-N
86.8% 86,8%
27.3%
88.1% 88,1%
25.9% 25,9%
6-N
86.2% 86,2%
28.5% 28,5%
88.5% 88,5%
26.3% 26,3%
ã È È ã È È
2"  )1
!   ãÈ 
  
 
 
.È ã

p)  

.È ã

p)  

1-U
95.1% 95,1%
83.1% 83,1%
90.5%
83.8% 83,8%
3-U
94.7% 94,7%
83.2% 83,2%
89.7% 89,7%
82.6% 82.6%
6-U
95.2% 95,2%
83.3% 83,3%
87.9%
82.6% 82.6%
1-P 1-P
95.6% 95,6%
83.0% 83,0%
89.5% 89,5%
83.7% 83,7%
3-P 3-P
95.7% 95,7%
82.7%
87.9%
82.9% 82,9%
6-P 6-P
95.06%
82.2%
90.6% 90,6%
82.0% 82,0%
1-N
46.5% 46,5%
96.5% 96,5%
45.3% 45,3%
96.4% 96,4%
3-N
44.6% 44,6%
93.4% 93,4%
39.6% 39,6%
93.3% 93,3%
6-N
45.5% 45,5%
92.9% 92,9%
37.4% 37,4%
92.8%

c  c 
Schrodt: Forecasting Conflict
Page 19 Page 19
From the perspective of forecasting, the N-type models are best at warning against the possible
"bolt out of the blue,". However, this comes at a price of a lot of false alarms: when an N-type
model forecasts a high-conflict week, there is only about a 30% chance that this will occur.
10 10
AA
high-conflict prediction by a U- or P-type model, meanwhile, will result in an actual high-
conflict
week in about 60% of the cases.
Time Series Analysis Time Series Analisis
Figures 4 through 6 show display the accuracy of the various models as a 5-week centered
moving average. bergerak rata-rata.
11 11
(In the event you are reading a photocopy of this, the original graphs are in
glorious color; these can be found in the electronic version of the paper. The correlation between
Korelasi antara
the three forecast periods is quite high, so you are not missing much in black-and-white.)
"Accuracy" is computed as the percentage of the "best" models that correctly classify each week.
For reference, Figure 7 shows the target sequence of WEIS 22-type events, at the same scale as
the
remaining figures (the heavy line is a 5-week centered moving average).
There are some general patterns to the accuracy. All three models are very accurate at the
beginning and end of the series, prior to about April 1992 and following April 1997. During the
Selama
period October-94 to October-97, the U/P and N models are almost perfect mirror images of each
other²when the accuracy of one type of model is high, the other is low, and vice-versa. (This
also
means that the models are making opposite predictions: for example following October-95, the P
models almost always [correctly] predict low conflict, whereas the N-models predict high
conflict.)
This is not merely a visual effect: the correlation (r) for the 3-N versus 3-P models is -0.59 for
Oct-91 to Mar-97, and -0.28 for the entire period.
Curiously, the N-model takes a very long time²almost eighteen months²to recover accuracy
after the implementation of the Dayton Accords, which might suggest that despite the fact that
these agreements resulted in a sudden cessation of hostilities, the preconditions for violence
remained in place and Dayton may in fact have had a major impact in actually enforcing peace.
The The
accuracy in the 1993 period is medium for both models; this is probably partly due to the fact
that
the high-low classifications are fluctuating rapidly during this period.
10 10
However, as noted below, some of these false positives are "near misses" in the sense of
predicting high conflict
in weeks that fall just short of the 20-event threshold. Furthermore, many of these errors occur
during the period
of the implementation of the Dayton Accords, and if that period is eliminated from the analysis,
the forecast
accuracy of the N models is probably closer to 50%.
11 11
The moving average is used for purposes of clarity²the weekly accuracy fluctuates very rapidly
in places,
resulting in a graph that is very difficult to read.

c  c 
Schrodt: Forecasting Conflict
Page 20 Page 20
Figures 8 and 9 [attempt to] show the accuracy for high-conflict weeks only for the P and N
models. model. These figures reinforce the points made earlier about the relative accuracy of the
two
approaches, and are primarily important for what they  show: (1) there is no clear
differentiation in the predictive accuracy of the 1, 3 and 6-month forecasts and (2) the accuracy
is
fairly consistent over time, except for the tails of the series (including 1992 for the N model).
p * 
È *
Accuracy of U-Models, 5-week Moving Average
00
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0.6 0.6
0.7 0.7
0.8 0.8
0.9 0.9
11
Jan.91
May.91
Sep.91
Jan.92
May.92
Sep.92
Jan.93
May.93
Sep.93
Jan.94
May.94
Sep.94
Jan.95
May.95
Sep.95
Jan.96
May.96
Sep.96
Jan.97
May.97
Sep.97
Jan.98
May.98
Sep.98 Sep.98
Jan.99 Jan.99
1-U
3-U
6-U

c  c 
Schrodt: Forecasting Conflict
Page 21 Page 21
p 3 
È 3
Accuracy of P-Models, 5-week Moving Average
00
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0.6 0.6
0.7 0.7
0.8 0.8
0.9 0.9
11
Jan.91
May.91
Sep.91
Jan.92
May.92
Sep.92
Jan.93
May.93
Sep.93
Jan.94
May.94
Sep.94
Jan.95
May.95
Sep.95
Jan.96
May.96
Sep.96
Jan.97
May.97
Sep.97
Jan.98
May.98
Sep.98 Sep.98
Jan.99 Jan.99
1-P 1-P
3-P 3-P
6-P 6-P

c  c 
Schrodt: Forecasting Conflict
Page 22
p 4 
È 4
Accuracy of N-Models, 5-week Moving Average
00
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0.6 0.6
0.7 0.7
0.8 0.8
0.9 0.9
11
Jan.91
May.91
Sep.91
Jan.92
May.92
Sep.92
Jan.93
May.93
Sep.93
Jan.94
May.94
Sep.94
Jan.95
May.95
Sep.95
Jan.96
May.96
Sep.96
Jan.97
May.97
Sep.97
Jan.98
May.98
Sep.98 Sep.98
Jan.99 Jan.99
1-N
3-N
6-N

c  c 
Schrodt: Forecasting Conflict
Page 23 Page 23
p % 
È %
Number of WEIS 22 Events per Week
00
10 10
20 20
30 30
40 40
50 50
60 60
70 70
80 80
90 90
100 100
Jan.91
May.91
Sep.91
Jan.92
May.92
Sep.92
Jan.93
May.93
Sep.93
Jan.94
May.94
Sep.94
Jan.95
May.95
Sep.95
Jan.96
May.96
Sep.96
Jan.97
May.97
Sep.97
Jan.98
May.98
Sep.98 Sep.98
Jan.99 Jan.99
WEIS 22
CMAV5

c  c 
Schrodt: Forecasting Conflict
Page 24 Page 24
p & 
È &
P-Model Accuracy, High Conflict Weeks Only
00
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0.6 0.6
0.7 0.7
0.8 0.8
0.9 0.9
11
Jan.91
May.91
Sep.91
Jan.92
May.92
Sep.92
Jan.93
May.93
Sep.93
Jan.94
May.94
Sep.94
Jan.95
May.95
Sep.95
Jan.96
May.96
Sep.96
Jan.97
May.97
Sep.97
Jan.98
May.98
Sep.98 Sep.98
Jan.99 Jan.99
1-P 1-P
3-P 3-P
6-P 6-P

c 
c 

Schrodt: Forecasting Conflict


Page 25
p + 
È +
N-Model Accuracy, High Conflict Weeks Only
00
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0.6 0.6
0.7 0.7
0.8 0.8
0.9 0.9
11
Jan.91
May.91
Sep.91
Jan.92
May.92
Sep.92
Jan.93
May.93
Sep.93
Jan.94
May.94
Sep.94
Jan.95
May.95
Sep.95
Jan.96
May.96
Sep.96
Jan.97
May.97
Sep.97
Jan.98
May.98
Sep.98 Sep.98
Jan.99 Jan.99
1-N
3-N
6-N

c  c 
Schrodt: Forecasting Conflict
Page 26 Page 26
( ( )  
All of the analyses reported above use templates from the entire period to develop the models.
This means that some of the "predictions" of the models²on average, 50% of the period²are in
fact retrospective in that they are based on templates that occur after the week that is being
classified. diklasifikasikan. In contrast, this section will evaluate two sets of models that are
purely predictive: the
templates will be chosen from the period prior to the weeks being evaluated.
Two different schemes were used to do this. The entire time period was subdivided by
calendar years: 1993, 1994, 1995, 1996, 1997 and 1998. Let C Biarkan C
kk
refer to the beginning of the first
full week of year k.
Prior Templates:
Templates were taken from any time prior to C
k; k;
predictions were evaluated
on all of the weeks greater than C
k. k.
Recent Templates: Templates were taken from the time period C
k-1 k-1
>t>C
k; k;
predictions were
evaluated on the weeks C
kk
>t>C
k+1 k +1
..
12 12
In other words, the ³prior´ scheme takes templates from any time before the beginning of a year,
and then evaluates the accuracy of the prediction on all of the remaining weeks in the data set,
whereas the ³recent´ scheme takes templates only from the previous year (where possible) and
evaluates accuracy on a single year. The relationship between these schemes is illustrated in
Figure 9. Gambar 9.
13 13
12 12
Because of the small number of high-conflict weeks at the beginning and end of the sequence,
there were not
enough high-conflict weeks to provide an adequate number of templates (8) for some years, so
this was
implements by chosing high-conflict templates from 1991 and 1992 for the year 1993, and for
1996 to the
beginning of the year for 1997 and 1998.
13 13
Note: Due to a last minute reorganization of the paper, there is a gap in the table and figure
numbers here.

c  c 
Schrodt: Forecasting Conflict
Page 27
p + 
È +
( )  )

Year C Tahun C
kk
Time period for templates
Time period for predictions
Full Penuh
Prior Sebelumnya
Recent Baru
These changes required relatively few changes in the program, but the estimation times increases
by roughly a factor of six: Using 16 Monte-Carlo experiments per year, the prior templates
estimation takes approximately 2000 minutes (33 hours) to run on a 350 Mhz Macintosh G3; the
recent templates estimation takes approximately 1400 minutes (23 hours). [The difference is due
to
the fact that the recent model needs to classify only 52 weeks, whereas the prior model evaluates
all
313 weeks]. Given these constraints, only the P and N models were evaluated.
The results of this analysis are reported in Tables 5 and 6. In general, both of the predictive
analyses mirror the full-sample analysis in the sense that most of the figures are within „ 20% of
the earlier results. sebelumnya hasil. The differences between the P- and N-models that were
found in the full-
sample analysis continue to be reflected in the predictive analysis, though it seems to be
somewhat
more attenuated. There also seems to be more of a tendency for the 6-month forecast to be less
accurate than the 1-month forecast, though these differences are frequently small (less than 10%)
and the pattern is not universal.
As expected, these results differ substantially over time, with the high-conflict years 1993, 1994
and 1995 generally having one pattern (consistently better or worse predictions, depending on
the
model), 1996 being an intermediate years, and 1997 and 1998 having a single pattern. There are
a Ada
few exceptions to this²for example the ³%High Forecast´ indicator for the P-model is
uniformly awful²but it holds more often than not. [Also note that the sample sizes in the later
c  c 
Schrodt: Forecasting Conflict
Page 28 Page 28
years of the prior analyses are relatively small²particularly for the high-conflict weeks²so the
accuracy fluctuates more wildly than it does in the full-sample analysis.]
It should be noted that in many cases, the purely-predictive models perform    than the full-
sample model, which is frequently not the case in statistical analyses. This is probably due to the
Hal ini mungkin disebabkan oleh
model being able to adapt to the changing characteristics of the system, for example the shift in
the
focus of the conflict from Croatia to Bosnia to Kosovo, as well as adapting to the periods of low
conflict. konflik. In some situations, this adaptation can be counter-productive, notably as the
prior model
adapts to the high-conflict period prior to 1996 and then finds almost no conflict to predict after
1996. 1996. This can lead to situations where there is a great deal of difference between the
relative
accuracy in predicting high-conflict and low-conflict weeks. But in a majority of the cases, the
short-term adaptation produces substantially better predictions. This will be discussed further in
Ini akan dibahas lebih lanjut dalam
the next section. bagian berikutnya.

c  c 
Schrodt: Forecasting Conflict
Page 29 Page 29
ã È 3  ã È 3  )) )( ( p) ã
5  ( )  6
)) )"
5 p (  7 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

53.57 53.57
68.80
74.15 74.15
78.96 78.96
90.16 90.16
94.31 94.31
''
60.73
61.86 61.86
61.38 61.38
81.21 81.21
82.75 82.75
90.85 90.85
44
67.33 67.33
59.39 59.39
67.34 67.34
73.29 73.29
89.41 89.41
85.71 85.71
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

39.55 39.55
27.93 27.93
20.83 20.83
35.00 35.00
12.50 12.50
6.25 6.25
''
29.11 29.11
31.52 31.52
35.19 35.19
33.75 33.75
31.25 31.25
0.00 0.00
44
23.54 23.54
48.54 48.54
29.63 29.63
40.00 40.00
25.00 25.00
25.00 25.00
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

57.76 57.76
77.61
81.89 81.89
80.37 80.37
90.89 90.89
95.91
''
70.18 70.18
68.41 68.41
65.19 65.19
82.73 82.73
83.24
92.50 92.50
44
80.43 80.43
61.73 61.73
72.82 72.82
74.36 74.36
90.01 90.01
86.82 86.82
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

21.89 21.89
21.19 21.19
14.31 14.31
5.41 5.41
1.27 1.27
2.70 2.70
''
22.61 22.61
17.70 17.70
12.79 12.79
5.90 5.90
1.71 1.71
0.00 0.00
44
26.47
21.47 21.47
13.66 13.66
4.76 4.76
2.29 2.29
3.33 3.33
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

76.16 76.16
83.32 83.32
87.69 87.69
97.47 97.47
99.11 99.11
98.25 98.25
''
76.79 76.79
82.25 82.25
87.39 87.39
97.50 97.50
99.23 99.23
98.07 98.07
44
77.86
84.76
87.70 87.70
97.48 97.48
99.23 99.23
98.45 98.45

c  c 
Schrodt: Forecasting Conflict
Page 30 Page 30
ã È 3È ã È 3È )) ) ( p) ã
5  ( )  6
)) )"
5 p (  7 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

53.86 53.86
40.24 40.24
37.35 37.35
28.14 28.14
42.42 42.42
69.53
''
55.93 55.93
26.30 26.30
26.73 26.73
23.10 23.10
30.32 30.32
56.70 56.70
44
67.82
32.08 32.08
25.29 25.29
21.16 21.16
27.84
48.55 48.55
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

44.26 44.26
84.18 84.18
82.87 82.87
96.25 96.25
50.00 50.00
37.50 37.50
''
37.59
91.22 91.22
90.05 90.05
83.75 83.75
56.25 56.25
37.50 37.50
44
23.20 23.20
89.63 89.63
93.75 93.75
93.75 93.75
87.50 87.50
56.25 56.25
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

56.74 56.74
30.76 30.76
30.75 30.75
25.96 25.96
42.35 42.35
70.11 70.11
''
61.42 61.42
12.30 12.30
17.54 17.54
21.15 21.15
30.08 30.08
57.05 57.05
44
81.17
19.67 19.67
15.36 15.36
18.83 18.83
27.28
48.41 48.41
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

23.44 23.44
20.77 20.77
14.80 14.80
4.00 4.00
0.80 0.80
2.23 2.23
''
22.57 22.57
18.32 18.32
13.68 13.68
3.29 3.29
0.75 0.75
1.56 1.56
44
26.94 26.94
19.39 19.39
13.85 13.85
3.57 3.57
1.11 1.11
1.94 1.94
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

77.29 77.29
90.02 90.02
92.52 92.52
99.54 99.54
98.91
98.41 98.41
''
76.69 76.69
86.67 86.67
92.39 92.39
97.60 97.60
98.66
98.05 98.05
44
77.94 77.94
89.79 89.79
94.42 94.42
98.95 98.95
99.57 99.57
98.38 98.38

c  c 
Schrodt: Forecasting Conflict
Page 31
ã È 4  ã È 4  )) )( ) ã
5 6( )   )) )È
7  ã 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

49.28 49.28
52.52 52.52
54.81 54.81
79.25 79.25
96.27 96.27
96.75 96.75
''
48.80 48.80
52.16 52.16
47.60 47.60
75.12 75.12
93.87 93.87
97.96 97.96
44
47.96 47.96
47.12 47.12
54.93 54.93
51.65
95.43 95.43
94.35 94.35
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

33.41 33.41
41.25 41.25
42.33 42.33
28.12 28.12
0.00 0.00
0.00 0.00
''
16.83 16.83
55.94 55.94
27.27 27.27
28.12 28.12
0.00 0.00
0.00 0.00
44
9.13 9.13
57.81 57.81
40.06 40.06
29.69 29.69
0.00 0.00
0.00 0.00
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

65.14
59.57
63.96 63.96
83.42 83.42
96.27 96.27
98.64
''
80.77 80.77
49.80 49.80
62.50 62.50
78.95
93.87 93.87
99.88 99.88
44
86.78 86.78
40.43 40.43
65.83 65.83
53.44 53.44
95.43 95.43
96.20 96.20
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

48.94 48.94
38.94 38.94
46.27
12.16 12.16
0.00 0.00
0.00 0.00
''
46.67 46.67
41.06 41.06
34.78 34.78
9.84 9.84
0.00 0.00
0.00 0.00
44
40.86 40.86
37.76 37.76
46.23
4.95 4.95
0.00 0.00
0.00 0.00
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

49.45 49.45
61.87 61.87
60.20 60.20
93.43 93.43
100.00 100.00
98.05 98.05
''
49.27 49.27
64.39 64.39
53.96 53.96
93.08 93.08
100.00 100.00
98.07 98.07
44
48.85 48.85
60.53 60.53
59.96 59.96
90.30 90.30
100.00 100.00
98.00 98.00

c  c 
Schrodt: Forecasting Conflict
Page 32 Page 32
ã È 4È ã È 4È )) ) ) ã
5 6( )   )) )È
7  ã 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

48.08 48.08
42.07 42.07
47.00 47.00
52.95 52.95
76.20 76.20
94.59 94.59
''
47.96 47.96
43.03 43.03
44.23 44.23
39.98 39.98
74.16 74.16
97.36 97.36
44
47.12 47.12
41.59 41.59
49.64 49.64
28.89 28.89
77.04 77.04
94.11 94.11
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

42.55 42.55
73.75 73.75
83.24
53.12 53.12
0.00 0.00
0.00 0.00
''
31.73 31.73
81.56 81.56
73.3 73.3
62.50 62.50
0.00 0.00
0.00 0.00
44
17.07 17.07
86.88 86.88
91.19 91.19
76.56 76.56
0.00 0.00
6.25 6.25
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

53.61 53.61
22.27 22.27
20.42 20.42
52.93 52.93
76.20 76.20
96.45 96.45
''
64.18 64.18
18.95 18.95
22.92 22.92
38.14 38.14
74.16 74.16
99.26 99.26
44
77.16 77.16
13.28 13.28
19.17 19.17
25.00 25.00
77.04 77.04
95.83 95.83
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

47.84 47.84
37.22 37.22
43.41 43.41
8.44 8.44
0.00 0.00
0.00 0.00
''
46.98 46.98
38.61 38.61
41.08 41.08
7.62 7.62
0.00 0.00
0.00 0.00
44
42.77 42.77
38.5 38.5
45.28 45.28
7.69 7.69
0.00 0.00
2.86 2.86
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

48.27 48.27
57.58 57.58
62.42 62.42
93.26
100.00 100.00
98.01 98.01
''
48.46 48.46
62.18 62.18
53.92 53.92
92.57 92.57
100.00 100.00
98.06 98.06
44
48.2 48.2
61.82 61.82
74.80 74.80
92.89 92.89
100.00 100.00
98.12 98.12

c  c 
Schrodt: Forecasting Conflict
Page 33 Page 33
Comparison of Forecast Accuracy by Year
The ³recent´ prediction scheme is the most sensitive to the changing character of the conflict.
The annual statistics in Table 6 are also the most straightforward to evaluate, because each
prediction is done on 52-weeks, in contrast to the variable number of weeks in the prediction
period in the ³prior´ scheme.
In order to get a systematic evaluation of the effect of changing the level of adaptation in the
model, the results from the full-sample and prior analysis were re-analyzed on an annual basis: In
other words, rather than assessing the accuracy on a period that extends to the end of the data set,
the accuracy for each year (52 weeks) of the data was tabulated. (The  of the model
was done as before²in fact this is simply a retabulation of the existing results, not a new set of
estimations.) These results are reported in Tables 7 and 8, which are directly comparable to
Table
6. 6.
Figures 10, 11 and 12 show the comparative level of accuracy of the three methods for the 3-
month lead. When the relative level of accuracy is compared across the estimation methods, two
patterns are evident in these figures.
First, there is generally a single rank ordering of the accuracy of the three methods across
time: For example, if the recent and prior schemes (or full-sample and recent schemes) have
roughly the same level of accuracy on one year, they will have this on all years. The exceptions
Pengecualian
that occur in this are usually at the ends of the data set, 1993 and 1998. There is quite a bit of
variation in the patterns across the accuracy measures and estimation techniques, however.
Second, the full-sample scheme is almost always better than the prior scheme; when this is no
true, the prior scheme and the recent scheme usually have about the same level of accuracy. In
all Dalam semua
cases, the prior scheme is either the least accurate of the three²in some cases dramatically less
accurate, as in Figure 12b , or else it is comparable in value to one of the other two measures.
These results suggest rather strongly that if one is in a predictive mode (as distinct from
retrospectively analyzing a period of time, or using one set of interactions to try to predict
behavior
in a different region), then it is best of use short-term adaptation. In many cases, the short-term
model does also as well as the full-sample model, and in Figure 10b, it does substantially better
in
the later years. HMMs seem to work best when they can ³forget´²in the sense of ignoring
older information²as well as ³learn.´
The optimal set of training templates might be one that effective has an ³exponential decay´ in
the sense of having more templates from recent history than from distant history. To a certain
Untuk sebuah tertentu
extent, this was done already in the latter years in this analysis, since it was necessary to go
further

c  c 
Schrodt: Forecasting Conflict
Page 34 Page 34
back in time to pick up sufficient examples of high-conflict weeks.
14 14
Simply accumulating
additional information on an on-going conflict as it becomes available²which is effectively
what
the ³prior´ scheme does²does not seem to be a good idea: from a statistical perspective, these
are non-stationary systems. On the other hand, one would still like to have some way of
maintaining a ³memory´ of the precursors to high-conflict situations in a situation which has
been peaceful for some time; this might be provided either by instances of earlier conflict in the
region (even if it occurred a number of years earlier), or with some generic templates for conflict
in
comparable regions. In Schrodt (1999) I was able to use hidden Markov models generated from a
set of 19th and 20th century crises (the BCOW data set) to provide a reasonably good measure of
conflict in the Israel-Palestine conflict, and based on this, it might be possible to find some good
archetypical models for conflict in areas that have generally been conflict-free.
14 14
However, this result has certainly been affected to some extent by the fact that the Balkans
conflict went
through four or five very distinct subphases²the initial conflict with Croatia, followed by the
conflict with
Bosnia, followed by the combined Croatia-Bosnian counter-offensive and brief NATO attacks,
followed by the
peaceful Dayton period, and ending with the lead-up to the Kosovo conflict. Because the coding
system is
sensitive to the presence or absence of activity by individual actors, it may to overly sensitive to
these shifts and
unable to pick up a general pattern for ³conflict.´ It would be interesting to compare these results
with a model
of a highly-institutionalized conflict such as Israel-Lebanon or Turkey-Kurds.

c 
c 

Schrodt: Forecasting Conflict


Page 35 Page 35
p  
È 
/")È7 ,'( 
00
20 20
40 40
60 60
80 80
100 100
120 120
1993 1993
1994 1994
1995 1995
1996 1996
1997 1997
1998 1998
FS-3-P
PR-3-P
RC-3-P
p È 
È È
/")È7 ,' 
00
20 20
40 40
60 60
80 80
100 100
120 120
1993 1993
1994 1994
1995 1995
1996 1996
1997 1997
1998 1998
FS-3-N
PR-3-N
RC-3-N

c  c 
Schrodt: Forecasting Conflict
Page 36
p 
/0 ")È7 ,'( 
00
10 10
20 20
30 30
40 40
50 50
60 60
1993 1993
1994 1994
1995 1995
1996 1996
1997 1997
1998 1998
FS-3-P
PR-3-P
RC-3-P
p È
/0 ")È7 ,' 
00
10 10
20 20
30 30
40 40
50 50
60 60
70 70
80 80
90 90
100 100
1993 1993
1994 1994
1995 1995
1996 1996
1997 1997
1998 1998
FS-3-N
PR-3-N
RC-3-N

c  c 
Schrodt: Forecasting Conflict
Page 37 Page 37
p  
È 
/2")È7 ,'( 
00
20 20
40 40
60 60
80 80
100 100
120 120
1993 1993
1994 1994
1995 1995
1996 1996
1997 1997
1998 1998
FS-3-P
PR-3-P
RC-3-P
p È 
È È
/2")È7 ,' 
00
20 20
40 40
60 60
80 80
100 100
120 120
1993 1993
1994 1994
1995 1995
1996 1996
1997 1997
1998 1998
FS-3-N
PR-3-N
RC-3-N

c  c 
Schrodt: Forecasting Conflict
Page 38 Page 38
ã È %  )) )( ( 
5 ã
5 6( ) 
)) )È7 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

48.44 48.44
51.08 51.08
52.88 52.88
67.45 67.45
95.31 95.31
94.11 94.11
''
48.80 48.80
49.88 49.88
47.36 47.36
76.30 76.30
88.34 88.34
91.71
44
47.00 47.00
49.52
53.37 53.37
65.57 65.57
93.39 93.39
87.74
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

39.18 39.18
35.31
19.03 19.03
34.38 ---
6.25 6.25
''
28.12 28.12
40.31 40.31
31.53
28.12 ---
00
44
23.32 23.32
56.56 56.56
29.26 29.26
39.06 ---
25.00 25.00
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

57.69 57.69
60.94 60.94
77.71
70.15 70.15
95.31 95.31
95.83 95.83
''
69.47 69.47
55.86 55.86
58.96
80.23 80.23
88.34 88.34
93.50 93.50
44
70.67
45.12 45.12
71.04 71.04
67.73 67.73
93.39 93.39
88.97 88.97
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

48.08 48.08
36.10 36.10
38.51 38.51
8.59 8.59
00
2.86 2.86
''
47.95 47.95
36.34 36.34
36.04 36.04
10.40 10.40
00
00
44
44.29 44.29
39.18 39.18
42.56 42.56
8.99 8.99
00
4.26 4.26
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

48.68
60.12
56.69 56.69
92.91 92.91
100 100
98.12 98.12
''
49.15 49.15
59.96 59.96
54.01 54.01
93.19 93.19
100 100
97.95 97.95
44
47.96 47.96
62.43 62.43
57.80 57.80
93.16 93.16
100 100
98.37 98.37

c  c 
Schrodt: Forecasting Conflict
Page 39 Page 39
ã È %È )) ) ( 
5 ã
5 6( ) 
)) )È7 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

46.15 46.15
39.30 39.30
42.91 42.91
14.98 14.98
31.01 31.01
69.95 69.95
''
47.24 47.24
40.75 40.75
41.95 41.95
17.45 17.45
16.47 16.47
56.49 56.49
44
46.27
42.43 42.43
45.67 45.67
17.57 17.57
23.56 23.56
48.56 48.56
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

41.35 41.35
87.19 87.19
83.81 83.81
98.44 ---
37.50 37.50
''
34.13 34.13
93.12 93.12
88.92 88.92
87.50 ---
37.50 37.50
44
24.76 24.76
90.62 90.62
94.60 94.60
93.75 ---
56.25 56.25
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

50.96 50.96
9.38 9.38
12.92 12.92
8.16
31.01 31.01
70.59 70.59
''
60.34 60.34
8.01 8.01
7.50 7.50
11.73 11.73
16.47 16.47
56.86 56.86
44
67.79 67.79
12.3 12.3
9.79 9.79
11.35 11.35
23.56 23.56
48.41 48.41
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

45.74 45.74
37.55
41.37 41.37
8.05 8.05
00
2.44 2.44
''
46.25 46.25
38.75 38.75
41.35 41.35
7.49 7.49
00
1.68 1.68
44
43.46 43.46
39.24 39.24
43.47 43.47
7.95 7.95
00
2.09 2.09
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

46.49 46.49
53.93 53.93
52.10 52.10
98.46 98.46
100 100
98.29
''
47.81 47.81
65.08
48.00 48.00
92.00 92.00
100 100
97.89 97.89
44
47.39 47.39
67.74 67.74
71.21 71.21
95.70 95.70
100 100
98.26 98.26

c  c 
Schrodt: Forecasting Conflict
Page 40 Page 40
ã È &  ã È &  )) )( p 
5 ã
5 6( )   )) )
È7   ã 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

56.32
65.11 65.11
62.64
90.03 90.03
100 100
97.53 97.53
''
54.55
64.16 64.16
66.61
90.57
97.90 97.90
97.38 97.38
44
62.18 62.18
64.74
62.18 62.18
87.42 87.42
98.72 98.72
95.51 95.51
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

25.56 25.56
25.18 25.18
20.27 20.27
14.29 ---
14.29 14.29
''
16.43 16.43
25.23 25.23
30.08 30.08
22.73 ---
9.09 9.09
44
32.89 32.89
25.00 25.00
16.67 16.67
0.00 ---
33.33 33.33
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

86.41
89.78 89.78
91.67 91.67
96.21 96.21
100.00 100.00
99.16 99.16
''
92.66 92.66
88.14 88.14
92.26 92.26
96.10 96.10
97.90 97.90
99.11 99.11
44
90.00 90.00
89.58 89.58
95.56 95.56
94.56 94.56
98.72 98.72
96.73 96.73
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

64.79 64.79
60.34 60.34
62.5 62.5
23.53 ---
25.00 25.00
''
69.12 69.12
56.70 56.70
73.20 73.20
32.26 32.26
00
16.67 16.67
44
75.76 75.76
60 60
73.33
00
00
16.67 16.67
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

54.27 54.27
66.01 66.01
62.66 62.66
93.22
100 100
98.33 98.33
''
52.58
65.68
65.26 65.26
93.84 93.84
100 100
98.23 98.23
44
58.54 58.54
65.65 65.65
60.99 60.99
92.05 92.05
100 100
98.67 98.67

c  c 
Schrodt: Forecasting Conflict
Page 41 Page 41
ã È &È ã È &È )) ) p 
5 ã
5 6( )   )) )
È7   ã 
/") /! 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

50.26 50.26
40.38 40.38
46.28 46.28
30.19 30.19
70.00
93.46
''
50.00 50.00
43.75 43.75
48.80 48.80
34.43 34.43
57.21 57.21
81.25 81.25
44
54.40 54.40
45.60 45.60
48.08 48.08
33.69 33.69
55.77 55.77
73.90 73.90
/0 ")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

95.86
97.79 97.79
94.70 94.70
95.00 ---
26.67 26.67
''
89.74 89.74
91.72 91.72
92.59 92.59
87.50 ---
50.00 50.00
44
92.44 92.44
92.91 92.91
90.14
89.29 ---
42.86 42.86
/2")
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

10.77 10.77
9.82 9.82
15.69 15.69
24.90
70.00
94.77 94.77
''
14.93
18.08 18.08
20.87 20.87
30.10 30.10
57.21 57.21
81.86
44
20.31 20.31
20.25 20.25
21.17 21.17
29.15 29.15
55.77 55.77
74.51 74.51
/0 p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

48.19 48.19
36.60 36.60
41.51 41.51
9.36 9.36
00
9.09 9.09
''
48.21 48.21
37.46 37.46
42.74 42.74
9.27 9.27
00
5.13 5.13
44
50.96 50.96
38.44 38.44
42.24 42.24
9.33 9.33
00
3.19 3.19
/2p) 
++' ++'
++* ++*
++3 ++3
++4 ++4
++% ++%
++& ++&

75.00 75.00
89.29
82.42 82.42
98.39
100 100
98.51 98.51
''
62.26 62.26
80.33 80.33
81.54 81.54
96.72 96.72
100 100
98.82 98.82
44
75.00 75.00
84.21 84.21
77.05 77.05
97.09 97.09
100 100
98.52 98.52

c  c 
Schrodt: Forecasting Conflict
Page 42 Page 42
Assessment of differences between the high-conflict
and low-conflict models
The HMM has a very diffuse coefficient structure: the high and low models each have 1,078
parameters (177 observation probabilities plus 16 transition probabilities), so determining how to
predict a week fits the high or low conflict category involves 2,156 parameters. Add the fact that
this analysis has looked at three different weighting schemes and three forecasting horizons, and
one is looking at 19,404 parameters; if the prospective and recent models are included, there are
58,212 parameters. Add the fact that we are using Monte Carlo methods, so that each of those
conditions was duplicated at least 16 times, and the analysis has actually produced in excess of
931,392 parameter estimates. That's a lot of parameters...
On the positive side, these parameters all have straightforward interpretations (in contrast, for
example, to the weights of a neural network). The larger observation probabilities correspond to
the behaviors that are being "watched" by the model in order to make a prediction. In this
analysis, Dalam analisis ini,
the transition probabilities are of interest primarily to determine the number of Markov states that
are actually found in the process. The   m  between the high and low models will show the
characteristics of the system that are most likely to distinguish between pre-conflict and pre-
peace
periods. periode.
In order to make sense of this mass of information, I'm focusing the analysis somewhat:
‡ Only the full-sample runs have been analyzed²these are likely to be most representative of
the results of the entire sample.
‡ Only the P and N models are compared²as in the prediction analysis, the parameter
estimates of the U and P models are very similar and with very few exceptions, anything true
of the P model will also be true of the N model.
‡ Most of the analysis will focus on the 3-month forecast²there are some differences between
the forecasting horizons, and these will be explored below²but in general the major
differences are found between the high and low conflict models, and between the P and N
weighting systems.
‡ Most of the analyses will be done using the "reduced" models (see discussion below)

c  c 
Schrodt: Forecasting Conflict
Page 43 Page 43
Because the parameter estimates result from Monte Carlo estimation, they will be analyzed
statistically. statistik. As before, the "All" results refer to the analysis of all of the Monte Carlo
runs; the
"Best" results are only those with accuracy ë 0.795 for P models and accuracy ë 0.495 for N
models. model.
State Reduction Criteria
This section gives details on the algorithm used to reduce the number of states in a model
through
elimination and combinion of states that will occur very rarely. In this discussion, the notations p
ii
,,
rr
ii
, and n , Dan n
ii
refer to the transitions to the previous state, same state, and next state respectively.
"Small" probabilities are²somewhat arbitrarily²defined as those less than 0.1; "large"
probabilities are defined as those greater than 0.9.
15 15
;;
Small values of r
ii
occur in two circumstances:
‡ transient states: r
ii
is small; p
ii
and p
i+1 i +1
are not large. tidak besar. In this case, the system will simply pass
through the state quickly and that state will be responsible for very few events in the sequence.
‡ cyclical pairs: r
ii
and r dan r
i+1 i +1
are small; n
ii
and p
i+1 i +1
are large. yang besar. In this case, the system will remain
inside this pair of states for a significant period of time, but switch between state i and i+1 for
every other event
While other combinations are logically possible²for example one could have a cyclical
triplet²virtually all of the cases of small r
ii
values fall into these categories.
The transient case can be simply eliminated from the model. The only situation where this
removes any important information is when p
ii
is very small, in which case the state is acting as a
one-way valve²once the system has passed through the state, it will not return to the earlier
states.
These situations are fairly common²in fact in a number of cases the estimated p
ii
is zero²but I
have not tried to analyze them separately.
The cyclical pair, in contrast, can be mathematically reduced to a single state. The pair forms a
Markov chain of the form
pp
rr
nn
pp
rr
nn
ii
i+1 i +1
i+1 i +1
i+1 i +1
ii
ii
Because r
ii
+n+N
ii
< 1 and p
i+1 i +1
+r
i+1 i +1
< 1, the middle matrix is not strictly a Markov chain, so these
are standardized by row, eg r
ii
is replaced with diganti dengan
rr
ii
rr
ii
+n+N
ii
..
A Markov chain relatively quickly reaches an equilibrium distribution where it spends a known
15 15
rr
ii
has a highly bimodel distribution...

c  c 
Schrodt: Forecasting Conflict
Page 44 Page 44
proportion of time in each of the states. Feller (1968: 432) shows that the Markov chain
pp
q1-q
1-p1-p
has the equilibrium distribution [
qq
p+qp+q
,,
pp
p+qp+q
]. ]. This formula will be used to give the weights to
the observation probabilities when reducing cyclical states. (In practice, n
ii
and p dan p
i+1 i +1
are sufficiently
close in value that this gives much the same results as simply weighting the two observation
vectors
equally)
Interpreting the t-test tables
Much of this analysis is based t-test tables such as the following:
HIGH CONFLICT
PARAMETER T-TESTS AGAINST MARGINALS FOR MODEL P3
State 0 Negara 0
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD -5.127
*** ***
--- ---
--- ---
+++ -3.948
--- ---
--- ---
COMMNT
--- ---
8.004 8.004
--- ---
2.495 2.321 12.233 -4.772 -3.849
CONSLT
--- ---
7.845 7.845
--- ---
--- ---
2.329 2.329
+++ -6.469 -8.034
APPROV
--- ---
-3.651
--- ---
+++ + + +
--- ---
-4.874
--- ---
-11.946
PROMIS
--- ---
--- ---
--- ---
--- ---
--- ---
-3.150 -3.150
--- ---
-3.085
GRANT GRANT
-2.735 -2.976
--- ---
--- ---
+++ + + +
--- ---
--- ---
-2.719
REWARD PENGHARGAAN
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
-4.266
AGREE SETUJU
-3.535 -3.592 -5.826 -7.120 -3.418 -2.324 -13.352 -10.067
REQEST
--- ---
+++ + + +
+++ -2.234 -2.466 -5.893
--- ---
-2.247
PROPOS
-6.552
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
-4.674
REJECT REJECT
--- ---
--- ---
+++ + + +
--- ---
3.707 3.707
+++ + + +
--- ---
-2.901
ACCUSE
+++ + + +
--- ---
--- ---
3.033 2.546
--- ---
--- ---
--- ---
PROTST
-2.252 -3.447
+++ + + +
--- ---
-2.396
--- ---
--- ---
-11.946
DENY +++ -3.255
--- ---
--- ---
2.502 -6.648
--- ---
+++ + + +
DEMAND
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
WARN WARN
-2.141 -2.141
--- ---
-2.679
*** ***
--- ---
-3.734 -3.734
--- ---
--- ---
THREAT ANCAMAN
-2.636
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
REDUCE MENGURANGI
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
-2.531
EXPEL ***
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
-8.707
SEIZE Merebut
--- ---
--- ---
--- ---
--- ---
*** ***
--- ---
--- ---
--- ---
FORCE FORCE
--- ---
-2.262 -3.914
--- ---
--- ---
-2.341
--- ---
--- ---
NULL ---
-2.135
--- ---
--- ---
--- ---
-3.501 -3.000 -8.602
NONEVT
-5.414
These tables are designed to make it relatively easy to determine where the significant relations
in the estimated observation probabilities exist. The first three lines²whose content will vary
with
the individual test²identify the test and the observation vector: In this case the vector is for
State 0
of the P weights, 3-month forecast, high-conflict model; the t-test compares the estimated
probabilities with the marginal probabilities.

c 
c 

Schrodt: Forecasting Conflict


Page 45 Page 45
The matrix of t-test has the event codes in the rows and the dyadic interactions *in the columns.
The rows are labeled with the WEIS codes (these are either mnemonics, as in this example, or
two-
digit numerical codes in some of the earlier runs). The columns are labeled with the dyads: the
notation BOS> refers to all actions where Bosnia is the  m ; >BOS refers to all actions where
Bosnia is the   (the notation <BOS is used in some earlier versions of the program). The
The
abbreviations are BOS = Bosnia; CRO = Croatia; SER = Serbia; KSV = Kosovo.
Only the significant t-tests are shown. The following thresholds are used:
|t | > 2.0
Print the actual value; this is approximately the critical value for the 0.05
significance level tingkat signifikansi
2.0 ë |t | > 1.72
Print "***" ; this is approximately the critical value for the 0.10 significance
level tingkat
1.72 ë |t | > 1.32 Print "+++"; this is approximately the critical value for 0.20 significance
level tingkat
1.32 ë |t | Print "---"
The selected critical values are for a t-test with 20 degrees-of-freedom. Note that the sample size
varies depending on the individual model, so these critical values are only approximate; in the
case
of some of the "best" tests, they slightly over-estimate the significance level.
16 16
When the reduced forms of the high and low models are being compared, there is an additional
problem of comparing the different number of models where a particular state exists (in other
words, because of state reduction, there will be fewer instances of models containing four, five or
six states than models having three states). There [obviously] must be at least two cases of a
particular state in both models (eg high and low; 1-month and 6-month) for the t-test to be
computed. dihitung.
Interpreting the t-test maps
In order to provide a means of visualizing all of the parameters simultaneously, this report
includes a number of color-coded maps that show the t-tests for all of the codes and dyads. The
The
16 16
It would be straightforward to adjust these tables so that they reflect true significance levels, but
since the levels
of significance tests are themselves rather arbitrary²contrary to popular belief in social science
statistical
circles, there is little evidence that Moses descended from Mt. Sinai bearing tablets that said "
Į €38
"²this
didn't seem to justify the trouble involved in entering at table into the program used to generate
these tables. In Dalam
all cases, the highest t-values are substantially greater than 2.0, and more generally, the number
of significant t-
values far exceeds what would be expected by chance under the null hypothesis of no
differences, so the critical
value of 2.0 is not particularly important in the analysis.

c  c 
Schrodt: Forecasting Conflict
Page 46 Page 46
red areas indicate t-test that have high positive values; the blue areas have high negative values;
and
the green areas are close to zero.
These maps are produced from surface-plots in Microsoft Excel, which introduces a few quirks.
First, the "points" being plotted are the    m  of the grid line, not the squares outlined by
the
grid (ideally, these plots should be done as a grid of appropriately colored squares, but this isn't
an
option in Excel). Excel also interpolates between adjacent points, so there may be several color
bands between two values. Technically, the codes are discrete so this is meaningless, though
because WEIS codes form a rough cooperation-to-conflict continuum, it makes some sense in the
horizontal direction (and probably enhances one's ability to make sense of the map). Second,
only
eight ranges are plotted, and in fact the "3 - 4" range is actually >3.0 for positive numbers and < -
3.0 for negatives. Finally, in order to keep the scaling consistent across the graphs, in a few cases
I
put an artificial "-3.0x" value in the KSV>:YIELD category.
Parameter comparisons
This section will discuss each of the following comparisons.
‡ Number of states in the reduced models;
‡ Pattern of coefficients that are significantly different from the observed distribution of the
codes in the data;
‡ Pattern of coefficients that are significantly different in the high and low models
‡ Comparison of 1-month and 6-month forecast models
The t-test tables themselves are very lengthy²each requires about 2500 lines, or roughly 40
single-spaced pages for the full model²so with a few exceptions these have not been included in
the paper. kertas. In some instances, summary-maps of the distributions will be used instead.

È  
 
As noted above, both transient and cyclic states occurred frequently in the estimates,
indicating²with one possible exception²that the six-state framework was adequate to account
for
the observed sequences. The tables that follow show the proportion of non-transient and cyclic
states in original matrices, and the average number of states in the reduced matrices, by weight
and
forecast period. periode peramalan.
As usual, the U/P and N models tend to mirror each other. In the P models, the high model
has about 20% cyclic states, while the low model has only about 1% to 2%; in the N model this
ratio is reversed. In all of the low models, only about 60% to 70% of the states in the low model

c  c 
Schrodt: Forecasting Conflict
Page 47
are non-transient (that is, neither transient nor cyclical); the high model in the U/P wieghts has
even
fewer non-transient state, but in the N model around 90% of the states are non-transient.
These characteristics are reflected in the average number of states in the reduced models. All
Semua
of the models except the N-weight high model average about 4.5 states. In most cases, the low
models have fewer states than the high models, though this difference is generally only around a
0.3 to 0.6 states. The standard deviations are relatively low.
17 17
The one exception is the N-type, high model, where the reduced models still have nearly six
states. negara. This would certainly suggest that moving to a larger model for this situation (or
for models
with similar weights) might be appropriate. This may be due to the fact that the N-type models
specialize in accurately forecasting the more complex, high-conflict category, which may require
the greater level of detail that is possible with a model with a greater number of states. P-type
models, in contrast, are better with low-conflict weeks, and there is a consistent tendency for
low-
conflict models to require fewer states than high-conflict models.
17 17
Because the distribution is bounded on the right²no model has more than six states²this is
probably not
normally distributed. terdistribusi normal. The standard deviations of the "best" models are
actually higher than those of the complete
set of models, which runs contrary to my expectations; this may be partly a function of the low
sample sizes.

c  c 
Schrodt: Forecasting Conflict
Page 48 Page 48
ã È + 
(5       )) )    
 )
  
 
Model Model
Total Jumlah
Non- Non-
Cyclic Siklik
%Non-
%Cyclic
states
transient sementara
transient sementara
U1 U1
HIGH:
132 132
69 69
25 25
52.27%
18.94% 18,94%
LOW:
132 132
91 91
11
68.94%
0.76% 0,76%
U3 U3
HIGH:
126 126
65 65
26 26
51.59% 51,59%
20.63%
LOW:
126 126
81 81
22
64.29% 64,29%
1.59% 1,59%
U6 U6
HIGH:
96 96
57 57
14 14
59.38%
14.58% 14,58%
LOW:
96 96
74 74
00
77.08%
0.00% 0.00%
P1 P1
HIGH:
174 174
93 93
37 37
53.45%
21.26% 21,26%
LOW:
174 174
115 115
88
66.09% 66,09%
4.60% 4,60%
P3 P3
HIGH:
324 324
184 184
69 69
56.79%
21.30% 21,30%
LOW:
324 324
204 204
55
62.96% 62,96%
1.54% 1,54%
P6 P6
HIGH:
96 96
55 55
18 18
57.29%
18.75% 18,75%
LOW:
96 96
72 72
11
75.00% 75,00%
1.04% 1,04%
N1 N1
HIGH:
96 96
87 87
11
90.62% 90,62%
1.04% 1,04%
LOW:
96 96
55 55
14 14
57.29%
14.58% 14,58%
N3 N3
HIGH:
96 96
92 92
11
95.83%
1.04% 1,04%
LOW:
96 96
51 51
17 17
53.12%
17.71% 17,71%
N6 N6
HIGH:
96 96
89 89
22
92.71%
2.08% 2,08%
LOW:
96 96
53 53
21 21
55.21% 55,21%
21.88% 21,88%

c  c 
Schrodt: Forecasting Conflict
Page 49 Page 49
ã È +È
(5       )) )    
 )
! . 
Model Model
Total Jumlah
Non- Non-
Cyclic Siklik
%Non-
%Cyclic
states negara
transient sementara
transient sementara
U1 U1
HIGH:
60 60
30 30
11 11
50.00% 50,00%
18.33% 18,33%
LOW:
60 60
38 38
11
63.33% 63,33%
1.67% 1,67%
U3 U3
HIGH:
42 42
21 21
99
50.00% 50,00%
21.43% 21,43%
LOW:
42 42
30 30
22
71.43%
4.76% 4,76%
U6 U6
HIGH:
12 12
88
11
66.67% 66,67%
8.33% 8,33%
LOW:
12 12
88
00
66.67% 66,67%
0.00% 0.00%
P1 P1
HIGH:
42 42
26 26
88
61.90% 61,90%
19.05%
LOW:
42 42
27 27
22
64.29% 64,29%
4.76% 4,76%
P3 P3
HIGH:
66 66
41 41
12 12
62.12%
18.18% 18,18%
LOW:
66 66
49 49
11
74.24% 74,24%
1.52% 1,52%
P6 P6
HIGH:
18 18
12 12
33
66.67% 66,67%
16.67% 16,67%
LOW:
18 18
12 12
00
66.67% 66,67%
0.00% 0.00%
N1 N1
HIGH:
90 90
81 81
11
90.00% 90,00%
1.11% 1,11%
LOW:
90 90
53 53
13 13
58.89%
14.44% 14,44%
N3 N3
HIGH:
48 48
46 46
00
95.83%
0.00% 0.00%
LOW:
48 48
27 27
77
56.25% 56,25%
14.58% 14,58%
N6 N6
HIGH:
42 42
37 37
11
88.10%
2.38% 2,38%
LOW:
42 42
22 22
10 10
52.38%
23.81%

c  c 
Schrodt: Forecasting Conflict
Page 50 Page 50
ã È  

È  ) 
  
 
Model Model
Total Jumlah
Non- Non-
Mean Berarti
Stdev STDEV
models transient per model
per model
U1 U1
HIGH:
22 22
94 94
4.27 4.27
0.686 0.686
LOW :
22 22
92 92
4.18 4.18
1.230 1.230
U3 U3
HIGH:
21 21
91 91
4.33
0.642 0.642
LOW :
21 21
83 83
3.95 3.95
0.898 0.898
U6 U6
HIGH:
16 16
71 71
4.44 4.44
0.788 0.788
LOW :
16 16
74 74
4.62 4.62
1.111 1.111
P1 P1
HIGH:
29 29
130 130
4.48 4.48
0.725 0.725
LOW :
29 29
122 122
4.21 4.21
1.349 1.349
P3 P3
HIGH:
54 54
249 249
4.61 4.61
0.591 0.591
LOW :
54 54
209 209
3.87 3.87
1.504 1.504
P6 P6
HIGH:
16 16
73 73
4.56 4.56
0.704 0.704
LOW :
16 16
73 73
4.56 4.56
0.998
N1 N1
HIGH:
16 16
88 88
5.50 5.50
1.061 1.061
LOW :
16 16
67 67
4.19 4.19
1.014 1.014
N3 N3
HIGH:
16 16
93 93
5.81 5.81
0.527 0.527
LOW :
16 16
68 68
4.25 4.25
0.750 0.750
N6 N6
HIGH:
16 16
91 91
5.69 5.69
0.583 0.583
LOW :
16 16
74 74
4.62 4.62
0.484 0.484

c  c 
Schrodt: Forecasting Conflict
Page 51 Page 51
ã È È

È  ) 
! . 
Model Model
Total Jumlah
Non- Non-
Mean Berarti
Stdev STDEV
models transient per model
per model
U1 U1
HIGH:
10 10
41 41
4.10 4.10
0.700 0.700
LOW :
10 10
39 39
3.90 3.90
1.136 1.136
U3 U3
HIGH:
77
30 30
4.29 4.29
0.700 0.700
LOW :
77
32 32
4.57 4.57
0.728 0.728
U6 U6
HIGH:
22
99
4.50 4.50
0.500 0.500
LOW :
22
88
4.00 4.00
1.000 1.000
P1 P1
HIGH:
77
34 34
4.86 4.86
0.350 0.350
LOW :
77
29 29
4.14 4.14
1.125 1.125
P3 P3
HIGH:
11 11
53 53
4.82 4.82
0.386 0.386
LOW :
11 11
50 50
4.55 4.55
0.782 0.782
P6 P6
HIGH:
33
15 15
5.00 5.00
-0.000
LOW :
33
12 12
4.00 4.00
0.816 0.816
N1 N1
HIGH:
15 15
82 82
5.47 5.47
1.087 1.087
LOW :
15 15
64 64
4.27 4.27
0.998
N3 N3
HIGH:
88
46 46
5.75 5.75
0.661 0.661
LOW :
88
34 34
4.25 4.25
0.829 0.829
N6 N6
HIGH:
77
38 38
5.43 5.43
0.728 0.728
LOW :
77
32 32
4.57 4.57
0.495 0.495

c  c 
Schrodt: Forecasting Conflict
Page 52 Page 52
"
5   È   5È È  
   È  )
   
The significant t-tests for some of the early states of the P3 and N3 models are shown in the
following tables and maps. In general, the observation probabilities in both the high and low
matrices differ substantially from the marginal (background) probabilities in the dataset as a
whole.
This is not surprising for the high model, but somewhat surprising for the low model, since about
80% of the weeks in the data are in the low-category. The fact that the low model also shows a
large number of differences is probably indicative of the fact that these models were selected to
maximize the   m between high and low forecast behaviors, so even the low model picks
up
distinctive patterns.
Keep in mind that because the observation probabilities must add to 1.0, disproportionately
high probabilities on some codes must be compensated by disproportionately low probabilities
on
some others; furthermore, those probabilities are summed across all of the dyads. Consequently
Karenanya
while a large positive t-value always means that the behavior is more common than would be
expected by chance, a large negative value can either mean that the behavior is being ignored
completely by the model, or simply that it does not receive emphasis proportional to that of the
high-frequency events (this, for example, is probably what is going on in the series of negative
values for Kosovo in the P3-LOW model, and Bosnia in the N3-LOW model). (The very large t-
values²those in excess of 10.0, or even in excess of 100.0²are usually associated with codes
that
occur only rarely and hence have standard deviations and means close to zero; these are usually
associated with Kosovo.)
Beyond this, several general points are evident from these tables. First, all of the models,
whether high or low, P or N, disproportionately pick up on WEIS "comment" and "consult"
codes
towards Bosnia (and, in some cases, Croatia), and usually from Serbia. Second, there is a Kedua,
ada
consistent pattern of the high models having a probability of non-events that is much lower than
expected by chance, where the low models have a non-event probability that is higher than
chance
(though usually with t-scores about 2 to 3, whereas the non-event t-scores in the high models are
in
the negative tens or hundreds). This reinforces what is becoming one of the fundamental of
event-
data-based early warning: the  m of a report is at least as important as the m   of the
report. laporan. Third, and unsurprisingly, violent behavior by Serbia²demonstrations, reduced
relations,
expulsions²is important in the high models, but not the low models.

c  c 
Schrodt: Forecasting Conflict
Page 53 Page 53
TABLE 11a.
HIGH CONFLICT PARAMETER T-TESTS AGAINST MARGINALS FOR MODEL P3
State 1 Negara 1
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
--- ---
--- ---
--- ---
-4.238
--- ---
+++ + + +
--- ---
--- ---
COMMNT
--- ---
8.235 8.235
--- ---
+++ + + +
*** 12.012 -3.440
--- ---
CONSLT
--- ---
7.563
--- ---
+++ 2.009 2.731
--- ---
+++ + + +
APPROV
--- ---
-4.286
--- ---
+++ + + +
--- ---
-5.017
--- ---
-3.614
PROMIS
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
GRANT GRANT
+++ + + +
*** ***
+++ + + +
--- ---
+++ + + +
--- ---
--- ---
--- ---
REWARD PENGHARGAAN
--- ---
--- ---
-2.993 -2.993
--- ---
--- ---
--- ---
--- ---
-3.429
AGREE SETUJU
-2.598
*** ***
+++ -2.042
*** -3.112 -8.630 -5.273
REQEST
--- ---
-2.792
--- ---
-2.085 -2.733 -3.798
--- ---
--- ---
PROPOS
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
REJECT REJECT
+++ + + +
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
ACCUSE
-9.528
--- ---
+++ -3.216
+++ -3.624 -540.692 -7.584
PROTST
--- ---
--- ---
--- ---
+++ + + +
--- ---
+++ + + +
--- ---
-3.639
DENY MENYANGKAL
--- ---
-5.058
+++ + + +
--- ---
--- ---
-2.934
*** ***
--- ---
DEMAND PERMINTAAN
--- ---
-3.096 -2.291
--- ---
--- ---
--- ---
--- ---
--- ---
WARN WARN
-2.089 -2.089
--- ---
--- ---
--- ---
*** -4.514
--- ---
--- ---
THREAT ANCAMAN
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
DEMONS
--- ---
--- ---
+++ + + +
--- ---
2.782 2.782
--- ---
--- ---
-2.246
REDUCE MENGURANGI
+++ + + +
--- ---
--- ---
--- ---
*** ***
--- ---
--- ---
-2.246
EXPEL
--- ---
2.311 2.311
--- ---
--- ---
*** ***
--- ---
--- ---
--- ---
SEIZE Merebut
-8.816 -2.981 -11.002 -2.422
--- ---
-8.630 -6.423
--- ---
FORCE FORCE
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
-2.019 -2.019
NONEVT
-41.565
State 2 Negara 2
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
COMMNT
--- ---
7.904
--- ---
--- ---
--- ---
10.426 -3.245 -2.345
CONSLT
--- ---
8.000 -2.955
--- ---
+++ + + +
--- ---
-2.779 -5.296
APPROV
--- ---
--- ---
--- ---
--- ---
--- ---
-3.263
--- ---
-6.024
PROMIS
--- ---
+++ + + +
+++ + + +
--- ---
*** ***
--- ---
--- ---
*** ***
GRANT GRANT
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
-5.419
--- ---
REWARD PENGHARGAAN
--- ---
--- ---
*** ***
--- ---
--- ---
+++ + + +
--- ---
-3.694
AGREE SETUJU
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
-8.745 -4.759
REQEST
--- ---
-4.392
*** ***
*** -2.466 -5.508
--- ---
--- ---
PROPOS
--- ---
--- ---
--- ---
--- ---
*** ***
+++ + + +
--- ---
-3.079 -3.079
REJECT REJECT
--- ---
-2.099
--- ---
-2.302
--- ---
-2.125 -2.125
--- ---
-3.416
ACCUSE
-2.589
--- ---
-3.751 -6.533 -4.144 -2.394
--- ---
-5.694
PROTST
*** ***
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
-6.014
DENY MENYANGKAL
--- ---
-2.279 -2.279
--- ---
--- ---
--- ---
-2.895
+++ + + +
--- ---
DEMAND PERMINTAAN
+++ -4.858
+++ + + +
--- ---
--- ---
-2.903
--- ---
--- ---
WARN WARN
--- ---
--- ---
--- ---
--- ---
2.157 2.157
--- ---
--- ---
--- ---
THREAT ANCAMAN
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
DEMONS
--- ---
--- ---
--- ---
--- ---
2.368 2.368
--- ---
--- ---
-4.265
REDUCE MENGURANGI
--- ---
+++ + + +
--- ---
+++ + + +
+++ + + +
--- ---
--- ---
-4.265
EXPEL
--- ---
+++ + + +
+++ -3.159
--- ---
*** ***
--- ---
--- ---
SEIZE Merebut
--- ---
-3.868 -11.873
+++ + + +
+++ -5.171
--- ---
-2.486
FORCE FORCE
--- ---
--- ---
*** ***
--- ---
+++ -3.516
*** -4.386
NONEVT
-41.449

c  c 
Schrodt: Forecasting Conflict
Page 54 Page 54
TABLE 11b.
LOW CONFLICT PARAMETER T-TESTS AGAINST MARGINALS FOR MODEL P3
State 1 Negara 1
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
-3.021
--- ---
--- ---
--- ---
+++ + + +
--- ---
-2.284 -2.110
COMMNT -3.305 10.613 2.106
--- ---
-3.004
9.739 9.739
*** -3.513
CONSLT
*** ***
9.396 9.396
--- ---
--- ---
--- ---
--- ---
*** -3.819
APPROV
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
-2.744 -2.744
PROMIS
-2.344 -2.992
--- ---
--- ---
*** ***
+++ -3.689 -3.730
GRANT GRANT
--- ---
--- ---
--- ---
--- ---
-4.142
+++ -3.875 -4.072
REWARD PENGHARGAAN
--- ---
-3.185 -2.208
2.410 2.410
--- ---
-2.269 -2.269
--- ---
-2.746
AGREE SETUJU
-2.804 -2.036
--- ---
--- ---
--- ---
-2.036 -5.106 -7.254
REQEST
-2.395
--- ---
+++ + + +
--- ---
*** ***
--- ---
--- ---
-3.731
PROPOS
-2.210 -4.489
*** ***
--- ---
-2.334 -4.089 -3.875 -3.569
REJECT REJECT
--- ---
--- ---
--- ---
+++ + + +
--- ---
*** -2.918 -2.986
ACCUSE
-2.065 -2.065
+++ 2.044
+++ + + +
*** ***
--- ---
--- ---
--- ---
PROTST
-2.785 -2.785
--- ---
2.339 2.339
--- ---
+++ 2.194
--- ---
--- ---
DENY MENYANGKAL
2.091 2.091
--- ---
--- ---
--- ---
-2.785 -2.785
--- ---
--- ---
-6.134
DEMAND PERMINTAAN
+++ -2.317
+++ + + +
*** ***
--- ---
+++ -3.011
--- ---
WARN WARN
--- ---
--- ---
-4.645
--- ---
-2.154 -2.096
--- ---
--- ---
THREAT ANCAMAN
-3.787
--- ---
2.334 2.251
--- ---
+++ + + +
*** -5.018
DEMONS
+++ + + +
--- ---
+++ 2.280
--- ---
--- ---
-4.318 -19.038
REDUCE MENGURANGI
*** -3.826
--- ---
-2.170
--- ---
--- ---
--- ---
--- ---
EXPEL
--- ---
--- ---
--- ---
--- ---
-2.245 -2.006
--- ---
--- ---
SEIZE Merebut
-2.293
--- ---
--- ---
2.501 -3.489
--- ---
--- ---
*** ***
FORCE FORCE
-8.186 -7.382
--- ---
+++ -3.163 -4.335
--- ---
-2.699
NONEVT
2.902 2.902
State 2 Negara 2
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
-2.686
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
COMMNT -5.039
4.700 4.700
--- ---
--- ---
--- ---
6.708
--- ---
--- ---
CONSLT
-3.229
5.660 5.660
--- ---
*** ***
--- ---
-3.625 -3.625
--- ---
--- ---
APPROV
-6.613 -4.744
--- ---
-3.211 -3.862 -2.880
--- ---
--- ---
PROMIS
--- ---
+++ -2.790
--- ---
--- ---
-4.696
--- ---
--- ---
GRANT GRANT
-8.145
+++ + + +
+++ + + +
+++ -7.440 -4.532
*** -3.407
REWARD PENGHARGAAN
-3.781
--- ---
-2.062 -2.062
--- ---
--- ---
--- ---
--- ---
--- ---
AGREE SETUJU
-2.036 -3.552 -2.022
--- ---
-2.725 -2.387 -6.433 -2.687
REQEST
*** ***
--- ---
--- ---
--- ---
--- ---
--- ---
-4.067
--- ---
PROPOS
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
REJECT REJECT
-4.314
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
ACCUSE
-6.935 -6.935
--- ---
--- ---
--- ---
-2.818 -2.818
--- ---
--- ---
--- ---
PROTST
-2.355
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
DENY MENYANGKAL
--- ---
-4.904
--- ---
*** -6.368 -2.382
--- ---
--- ---
DEMAND -9.357
--- ---
*** ***
--- ---
--- ---
-2.876
--- ---
--- ---
WARN WARN
-45.513
--- ---
--- ---
--- ---
-3.054
--- ---
+++ + + +
--- ---
THREAT ANCAMAN
-8.092 -5.045
--- ---
--- ---
-3.503
+++ + + +
--- ---
--- ---
DEMONS
-3.597 -3.133
--- ---
+++ + + +
+++ -6.083
--- ---
-5.202
REDUCE MENGURANGI
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
EXPEL
--- ---
-2.126
--- ---
--- ---
-2.691 -2.326 -43.497 ---
SEIZE Merebut
-11.142 -2.142
--- ---
--- ---
-3.741
--- ---
--- ---
--- ---
FORCE FORCE
-2.038 -2.700
--- ---
--- ---
-2.911 -2.719
--- ---
--- ---
NONEVT
1.993 1.993

c 
c 

Schrodt: Forecasting Conflict


Page 55 Page 55
TABLE 11c.
HIGH CONFLICT PARAMETER T-TESTS AGAINST MARGINALS FOR MODEL N3
State 1 Negara 1
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
--- ---
--- ---
--- ---
-2.471 -2.471
--- ---
--- ---
--- ---
--- ---
COMMNT
--- ---
7.624 7.624
--- ---
--- ---
*** ***
9.578
*** -33.386
CONSLT
--- ---
4.392 4.392
--- ---
--- ---
--- ---
--- ---
-10.062 -20.709
APPROV
--- ---
--- ---
--- ---
--- ---
--- ---
-5.779
--- ---
--- ---
PROMIS
-5.755
--- ---
--- ---
+++ -2.907
--- ---
--- ---
-81.591
GRANT GRANT
--- ---
--- ---
--- ---
--- ---
2.504
--- ---
--- ---
+++ + + +
REWARD PENGHARGAAN
+++ + + +
--- ---
--- ---
-3.524 -3.524
--- ---
+++ + + +
--- ---
-98.116
AGREE SETUJU
--- ---
--- ---
--- ---
-2.325
--- ---
-2.612 -41.931 -20.400
REQEST
--- ---
--- ---
--- ---
--- ---
-5.145
--- ---
--- ---
--- ---
PROPOS
--- ---
--- ---
--- ---
--- ---
*** ***
--- ---
--- ---
-68.371
REJECT REJECT
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
ACCUSE
--- ---
--- ---
-6.945
--- ---
-2.089 -2.089
--- ---
+++ + + +
--- ---
PROTST
--- ---
*** ***
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
DENY MENYANGKAL
--- ---
-7.021 -2.367 -32.017 -2.243 -5.772
--- ---
--- ---
DEMAND PERMINTAAN
--- ---
--- ---
+++ + + +
--- ---
-2.795
--- ---
--- ---
--- ---
WARN WARN
-3.000 -6.464
--- ---
+++ + + +
--- ---
+++ + + +
--- ---
--- ---
THREAT ANCAMAN
--- ---
--- ---
2.861 2.861
*** ***
--- ---
--- ---
--- ---
--- ---
DEMONS
--- ---
-2.641 -2.641
*** ***
--- ---
+++ + + +
--- ---
--- ---
-147.690
REDUCE MENGURANGI
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
EXPEL
--- ---
--- ---
+++ + + +
--- ---
--- ---
+++ + + +
--- ---
--- ---
SEIZE Merebut
--- ---
--- ---
-6.470 -2.171
--- ---
-3.422 -3.088 -6.566
FORCE FORCE
--- ---
--- ---
*** ***
--- ---
--- ---
-4.190 -14.170 -5.912
NONEVT
-19.223
State 2 Negara 2
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
--- ---
--- ---
-2.061 -2.775
--- ---
--- ---
--- ---
--- ---
COMMNT
--- ---
6.998 6.998
--- ---
--- ---
+++ 7.518 -81.591 -18.838
CONSLT
--- ---
6.145
--- ---
--- ---
+++ + + +
*** -10.333 -10.953
APPROV
--- ---
-2.859 -2.335
--- ---
--- ---
-2.243 -2.243
--- ---
--- ---
PROMIS
-2.125 -2.125
--- ---
--- ---
*** ***
--- ---
--- ---
--- ---
--- ---
GRANT GRANT
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
REWARD PENGHARGAAN
+++ + + +
*** ***
--- ---
--- ---
--- ---
+++ + + +
--- ---
-3.828
AGREE SETUJU
+++ + + +
--- ---
--- ---
-2.048
--- ---
-3.019 -14.131
*** ***
REQEST
--- ---
--- ---
+++ + + +
--- ---
-2.579 -3.114
--- ---
-3.474
PROPOS
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
-4.552
REJECT REJECT
--- ---
--- ---
+++ + + +
--- ---
+++ + + +
--- ---
--- ---
-3.349
ACCUSE
+++ + + +
--- ---
-2.619
+++ -3.834 -2.413 -81.591 ---
PROTST
--- ---
--- ---
--- ---
--- ---
2.169 2.169
--- ---
--- ---
--- ---
DENY MENYANGKAL
+++ -3.073
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
DEMAND PERMINTAAN
--- ---
--- ---
--- ---
--- ---
-2.445
--- ---
--- ---
--- ---
WARN WARN
--- ---
--- ---
--- ---
+++ + + +
--- ---
-2.001
--- ---
--- ---
THREAT ANCAMAN
--- ---
--- ---
+++ + + +
*** ***
+++ + + +
+++ + + +
--- ---
--- ---
DEMONS
-2.412 -2.412
--- ---
--- ---
*** ***
--- ---
-2.100 -2.100
--- ---
-36.148
REDUCE MENGURANGI
+++ + + +
--- ---
--- ---
--- ---
2.635
--- ---
--- ---
--- ---
EXPEL
--- ---
+++ + + +
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
SEIZE Merebut
--- ---
-2.733 -8.688
+++ + + +
--- ---
-10.030 -2.066 -4.113
FORCE FORCE
--- ---
--- ---
-2.062 -2.062
--- ---
--- ---
-2.662 -6.771 -10.227
NONEVT
-8.849

c  c 
Schrodt: Forecasting Conflict
Page 56 Page 56
TABLE 11d.
LOW CONFLICT PARAMETER T-TESTS AGAINST MARGINALS FOR MODEL N3
State 1 Negara 1
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
-2.319
*** -2.059
--- ---
--- ---
*** ***
--- ---
--- ---
COMMNT
--- ---
3.921 3.921
--- ---
*** ***
--- ---
6.832 6.832
--- ---
--- ---
CONSLT
-2.326 -2.326
4.006
--- ---
--- ---
--- ---
--- ---
--- ---
*** ***
APPROV
-2.238
--- ---
+++ -4.024
+++ + + +
--- ---
--- ---
--- ---
PROMIS
--- ---
+++ + + +
--- ---
--- ---
--- ---
-2.332
--- ---
--- ---
GRANT GRANT
--- ---
-6.053 -6.053
--- ---
--- ---
-9.946
--- ---
--- ---
--- ---
REWARD -11.594 -8.933
--- ---
+++ -3.205
--- ---
--- ---
--- ---
AGREE SETUJU
-4.863 -5.890
--- ---
--- ---
*** ***
--- ---
-52.673 ---
REQEST
-2.734 -2.734
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
PROPOS
--- ---
-4.190
--- ---
--- ---
-2.441
--- ---
--- ---
-3.804
REJECT REJECT
--- ---
--- ---
--- ---
--- ---
2.043 2.043
*** ***
--- ---
--- ---
ACCUSE
-7.449
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
PROTST
-2.150
*** ***
--- ---
--- ---
--- ---
+++ + + +
--- ---
+++ + + +
DENY MENYANGKAL
-199.153
-3.634 -3.634
--- ---
--- ---
*** ***
--- ---
--- ---
--- ---
DEMAND -7.957
*** ***
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
WARN WARN
--- ---
-4.334
*** ***
--- ---
--- ---
-3.851
--- ---
--- ---
THREAT ANCAMAN
-4.137
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
DEMONS
-2.660 -3.320 -2.246
--- ---
--- ---
--- ---
--- ---
--- ---
REDUCE MENGURANGI
--- ---
-5.902
--- ---
--- ---
+++ + + +
--- ---
--- ---
*** ***
EXPEL
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
SEIZE Merebut
+++ + + +
--- ---
-2.795
--- ---
--- ---
*** ***
--- ---
--- ---
FORCE FORCE
-3.540 -3.230
*** ***
--- ---
--- ---
--- ---
--- ---
--- ---
NONEVT
1.887 1.887
State 2 Negara 2
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
-3.791 -4.067
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
COMMNT -4.534
3.671 3.671
--- ---
--- ---
-4.211 -4.211
3.509 3.509
--- ---
-17.856
CONSLT
-3.198
4.490 4.490
+++ + + +
--- ---
--- ---
--- ---
-9.154 -8.159
APPROV
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
PROMIS
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
-4.416 -4.416
GRANT GRANT
-4.189 -3.838
--- ---
2.529 -4.810
--- ---
--- ---
--- ---
REWARD PENGHARGAAN
-7.994 -2.711
--- ---
--- ---
+++ -4.244
--- ---
--- ---
AGREE SETUJU
--- ---
--- ---
*** ***
*** ***
*** ***
2.180 2.180
--- ---
-4.656
REQEST
-3.705 -3.318
--- ---
--- ---
+++ + + +
--- ---
-131.165
-14.831
PROPOS
*** -2.498
--- ---
--- ---
-7.115 -3.055
--- ---
-2.085
REJECT REJECT
--- ---
-2.622
--- ---
--- ---
-3.241 -3.251
--- ---
-131.165
ACCUSE
--- ---
-2.413
*** ***
*** ***
--- ---
--- ---
-4.718
--- ---
PROTST
-2.601
--- ---
--- ---
--- ---
-5.090
--- ---
--- ---
--- ---
DENY MENYANGKAL
--- ---
--- ---
--- ---
2.079
--- ---
+++ + + +
--- ---
--- ---
DEMAND -7.780 -3.605
--- ---
--- ---
--- ---
--- ---
-40.279 ---
WARN WARN
--- ---
+++ + + +
--- ---
--- ---
-2.512
--- ---
--- ---
--- ---
THREAT ANCAMAN
+++ + + +
--- ---
--- ---
2.043 2.043
+++ -2.836
--- ---
-73.328
DEMONS
--- ---
-5.320
--- ---
--- ---
--- ---
--- ---
--- ---
*** ***
REDUCE MENGURANGI
--- ---
-4.059
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
EXPEL
--- ---
--- ---
*** ***
--- ---
-2.226 -32.017
--- ---
--- ---
SEIZE Merebut
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
FORCE FORCE
-9.893 -6.976
--- ---
--- ---
+++ + + +
*** ***
--- ---
-4.468 -4.468
NONEVT
0.908 0.908

c  c 
Schrodt: Forecasting Conflict
Page 57 Page 57
p '
#9pp$$"$.p$ ã$ã!$ã1$$$ã9 ã$# #
 9 (.! !99ã9$,'.ã0090.#$ã ã$
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
('090ã ã$
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4
YIELD HASIL
COMMNT
CONSLT
APPROV
PROMIS
GRANT GRANT
REWARD PENGHARGAAN
AGREE SETUJU
REQEST
PROPOS
REJECT REJECT
ACCUSE
PROTST
DENY MENYANGKAL
DEMAND PERMINTAAN
WARN WARN
THREAT ANCAMAN
DEMONS
REDUCE MENGURANGI
EXPEL
SEIZE Merebut
FORCE FORCE
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
'090ã ã$
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4

c  c 
Schrodt: Forecasting Conflict
Page 58 Page 58
p 'È
#9pp$$"$.p$ ã$ã!$ã1$$$ã9 ã$# #
 9 (.! !99ã9$,'.ã0.1.#$ã ã$
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
('.1ã ã$
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
'.1ã ã$
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4

c   c  
Schrodt: Forecasting Conflict
Page 59 Page 59
#  )È2    2
 
T±tests were run for the differences between the high-conflict and low-conflict forecasting
models. model. The differences between the observation probabilities in the high and low models
generally show more consistent patterns than the parameter values themselves. The tables and
maps show the difference of means of the high model minus the low model, so positive values
indicate that the high model is higher.
The most obvious feature of both the tables and the maps is that almost all of the values are
positive. positif. This is an effect of the much lower number of non-events in the periods
preceding a
high-conflict period. In the N3 model, most of the differences involve Bosnia and Kosovo in the
state 1 model, and in the state 4 model, behavior towards Kosovo²cooperative and
conflictual²dominates. These strong results on Kosovo are undoubtedly due in part to the small
standard deviations, but emphasis on cooperative events may also reflect attention by the
international community towards Kosovo when tensions in the region heat up. Serbian
reductions
of relations and expulsions are picked up disproportionately in the state 1, but not the state 4,
models, and curiously so are grants and rewards. In state 4, the high model looks at agreements
across all of the dyads²this probably reflects international attempts to mediate. Interestingly, the
Menariknya,
maps for all models and the best models are almost identical.
The P3 model is quite different (and, in contrast to the N3 model, the all and best models are
different). berbeda). This model puts far more emphasis on violent behavior that is coming from
Serbia, and
Kosovo is generally ignored in the state 1 model. The state 4 matrix, in contrast, tends to look
similar to the patterns found in the state 4 matrix of the N3 model, albeit with substantially more
emphasis on Bosnian behavior.
The tables also report t-tests on the differences between the Markov probabilities in the
models. model. Here one finds a consistent pattern: in the P-type models the recurrence
probabilities are
consistently (and often dramatically) higher in the high model than in the low model; the
opposite
is just as consistently true for the N-type models. In the case of the N-type models, this may
simply reflect the tendency of the low model to have fewer states than the high model, but that
also
holds for P-type models, albeit in general the P-type models have fewer states than N-type
models.
The statistics in the tables as a whole (including those not presented here) provides some
evidence,
in the form of much lower prior transition probabilities, that the P-type high model is generally
moving forward, state by state, whereas the low model simply fluctuates forward and backwards
between several different states. This would be consistent with the high model showing
escalatory
behavior while the low model handles a background of fluctuating behavior. In the N-type
model,
this pattern is reversed: the high model fluctuates, whereas the low model shows de-escalation.

c   c  
Schrodt: Forecasting Conflict
Page 60 Page 60
TABLE 12a.
HIGH .VS. LOW T-TESTS FOR MODEL P3
State 0 Negara 0
Prev_Trans
NaN NaN
Same_Trans
3.224 3.224
Next_Trans
-3.224
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
-2.562 -3.526
COMMNT
3.122
--- ---
+++ + + +
*** ***
3.638 3.638
--- ---
+++ + + +
*** ***
CONSLT
*** ***
--- ---
--- ---
--- ---
2.246 2.350 -3.653 -2.638
APPROV
--- ---
-2.313
--- ---
+++ + + +
--- ---
-4.218 -3.161 -2.972
PROMIS
--- ---
--- ---
*** ***
+++ + + +
--- ---
-2.335
+++ + + +
+++ + + +
GRANT GRANT
-2.899 -2.898
--- ---
+++ 3.576
--- ---
*** ***
--- ---
REWARD PENGHARGAAN
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
-3.552
*** ***
AGREE SETUJU
--- ---
+++ -4.179 -4.162
*** ***
*** -2.406 -3.304
REQEST
--- ---
--- ---
-2.521 -2.360 -2.498 -3.361
--- ---
--- ---
PROPOS
-2.066
--- ---
+++ + + +
+++ + + +
*** ***
3.143 3.143
--- ---
-2.547
REJECT REJECT
+++ + + +
+++ + + +
--- ---
+++ 3.196
--- ---
-3.148 -2.087
ACCUSE
--- ---
--- ---
--- ---
2.043 3.196
--- ---
--- ---
--- ---
PROTST
--- ---
-2.677
--- ---
--- ---
*** ***
--- ---
-4.285 -3.109
DENY MENYANGKAL
*** -2.002 -2.421
--- ---
2.403 -5.277
--- ---
+++ + + +
DEMAND PERMINTAAN
+++ + + +
--- ---
*** ***
--- ---
--- ---
--- ---
-2.931 -3.675
WARN WARN
--- ---
--- ---
--- ---
+++ 2.294
+++ + + +
+++ -2.305
THREAT ANCAMAN
--- ---
--- ---
-2.040
--- ---
--- ---
*** -2.981 -3.539
DEMONS
--- ---
--- ---
--- ---
-2.462
--- ---
--- ---
--- ---
--- ---
REDUCE MENGURANGI
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
-2.774 -2.774
EXPEL
2.063 2.063
--- ---
--- ---
+++ 2.332
--- ---
--- ---
--- ---
SEIZE Merebut
--- ---
--- ---
+++ + + +
*** ***
2.308 2.308
--- ---
--- ---
--- ---
FORCE FORCE
3.293 3.293
+++ + + +
--- ---
--- ---
3.176 3.176
--- ---
-3.339 -3.475
NONEVT
-4.815
State 1 Negara 1
Prev_Trans -10.373
Same_Trans
9.481 9.481
Next_Trans
-1.410 -1.410
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
--- ---
--- ---
--- ---
-2.979
--- ---
--- ---
-2.515 -3.784
COMMNT
+++ + + +
+++ -2.064
+++ 2.952
--- ---
--- ---
--- ---
CONSLT
*** ***
--- ---
--- ---
--- ---
+++ 2.532
--- ---
--- ---
APPROV
--- ---
-2.671
--- ---
+++ + + +
--- ---
-2.594 -2.539
--- ---
PROMIS
--- ---
*** ***
--- ---
--- ---
+++ + + +
--- ---
+++ + + +
--- ---
GRANT GRANT
--- ---
--- ---
--- ---
--- ---
2.865 2.865
--- ---
--- ---
2.149
REWARD PENGHARGAAN
--- ---
*** ***
--- ---
-2.028
--- ---
+++ -2.807
--- ---
AGREE SETUJU
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
REQEST
--- ---
-2.028
+++ + + +
+++ -3.284 -2.494
--- ---
--- ---
PROPOS
+++ 2.316
--- ---
+++ + + +
+++ + + +
--- ---
--- ---
--- ---
REJECT REJECT
+++ + + +
--- ---
--- ---
*** ***
--- ---
--- ---
-2.030
--- ---
ACCUSE
-3.522
--- ---
-2.545 -2.938
--- ---
-3.033 -3.078 -3.003
PROTST
--- ---
--- ---
+++ + + +
--- ---
+++ -2.668 -3.676
+++ + + +
DENY MENYANGKAL
+++ -2.138
--- ---
--- ---
+++ -2.404
+++ + + +
+++ + + +
DEMAND PERMINTAAN
--- ---
--- ---
-2.322
+++ + + +
--- ---
--- ---
-2.220 -3.282
WARN WARN
--- ---
--- ---
--- ---
--- ---
2.630 2.630
+++ + + +
*** -2.339
THREAT ANCAMAN
*** ***
--- ---
*** ***
--- ---
--- ---
--- ---
-3.106 -2.690
DEMONS
*** ***
--- ---
--- ---
-2.154
2.365
--- ---
--- ---
+++ + + +
REDUCE MENGURANGI
2.066 2.066
*** ***
--- ---
--- ---
*** ***
--- ---
--- ---
+++ + + +
EXPEL
--- ---
2.188
--- ---
--- ---
2.150 2.150
+++ + + +
--- ---
--- ---
SEIZE Merebut
-2.717 -2.639 -4.498 -3.393
2.065 -4.378 -2.651
--- ---
FORCE FORCE
4.043 2.737
--- ---
--- ---
2.122 2.122
--- ---
--- ---
--- ---
NONEVT
-4.185

c   c  
Schrodt: Forecasting Conflict
Page 61 Page 61
TABLE 12b.
HIGH .VS. LOW T-TESTS FOR MODEL N3
State 0 Negara 0
Prev_Trans
NaN NaN
Same_Trans
-3.352
Next_Trans
3.352 3.352
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
--- ---
+++ + + +
--- ---
--- ---
2.253 2.253
--- ---
*** ***
--- ---
COMMNT
--- ---
--- ---
--- ---
--- ---
--- ---
-2.622
--- ---
--- ---
CONSLT
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
*** -2.016
APPROV
--- ---
--- ---
--- ---
--- ---
--- ---
-2.594
--- ---
-2.175
PROMIS
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
GRANT GRANT
--- ---
--- ---
2.027 2.027
*** ***
*** ***
--- ---
--- ---
--- ---
REWARD PENGHARGAAN
--- ---
--- ---
--- ---
--- ---
-2.051 -2.051
--- ---
--- ---
-2.064
AGREE SETUJU
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
REQEST
*** ***
--- ---
--- ---
+++ + + +
+++ + + +
*** ***
--- ---
-2.723 -2.723
PROPOS
--- ---
-2.259 -2.259
+++ + + +
--- ---
+++ + + +
+++ + + +
--- ---
+++ + + +
REJECT REJECT
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
+++ -2.051
ACCUSE
--- ---
*** ***
--- ---
--- ---
2.193 2.193
--- ---
+++ + + +
+++ + + +
PROTST
--- ---
+++ + + +
--- ---
+++ + + +
+++ + + +
--- ---
--- ---
--- ---
DENY MENYANGKAL
*** ***
--- ---
--- ---
--- ---
--- ---
-3.119
--- ---
--- ---
DEMAND PERMINTAAN
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
*** ***
*** ***
WARN WARN
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
THREAT ANCAMAN
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
DEMONS
+++ -2.259
--- ---
*** ***
--- ---
--- ---
--- ---
--- ---
REDUCE MENGURANGI
--- ---
*** ***
--- ---
--- ---
--- ---
--- ---
*** ***
+++ + + +
EXPEL
--- ---
*** ***
--- ---
*** ***
3.409 3.409
--- ---
--- ---
--- ---
SEIZE Merebut
--- ---
--- ---
*** ***
--- ---
--- ---
--- ---
--- ---
--- ---
FORCE FORCE
2.239 2.239
--- ---
--- ---
--- ---
--- ---
--- ---
+++ -3.112
NONEVT
-2.402
State 1 Negara 1
Prev_Trans
3.853 3.853
Same_Trans
-5.535
Next_Trans
4.039 4.039
BOS> >BOS CRO> >CRO SER> >SER KSV> >KSV
YIELD HASIL
--- ---
+++ + + +
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
COMMNT
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
+++ + + +
CONSLT
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
+++ + + +
APPROV
+++ + + +
--- ---
+++ + + +
--- ---
--- ---
+++ + + +
--- ---
+++ + + +
PROMIS
+++ + + +
+++ + + +
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
GRANT GRANT
--- ---
2.079
--- ---
--- ---
4.514 4.514
--- ---
--- ---
--- ---
REWARD PENGHARGAAN
3.516 3.680
--- ---
--- ---
*** ***
--- ---
+++ + + +
+++ + + +
AGREE SETUJU
+++ + + +
*** ***
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
REQEST
+++ + + +
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
+++ + + +
PROPOS
--- ---
2.575 2.575
+++ + + +
--- ---
2.623
--- ---
--- ---
+++ + + +
REJECT REJECT
--- ---
--- ---
--- ---
+++ + + +
--- ---
*** ***
--- ---
--- ---
ACCUSE
2.450 2.450
--- ---
-2.098 -2.098
--- ---
--- ---
--- ---
--- ---
+++ + + +
PROTST
--- ---
--- ---
+++ + + +
--- ---
+++ + + +
--- ---
*** ***
+++ + + +
DENY MENYANGKAL
+++ + + +
--- ---
--- ---
--- ---
--- ---
*** ***
--- ---
--- ---
DEMAND PERMINTAAN
2.157 2.157
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
WARN WARN
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
--- ---
THREAT ANCAMAN
+++ + + +
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
DEMONS
--- ---
--- ---
2.420
+++ + + +
--- ---
--- ---
--- ---
+++ + + +
REDUCE MENGURANGI
--- ---
2.646 2.646
--- ---
--- ---
*** ***
--- ---
--- ---
--- ---
EXPEL
--- ---
--- ---
--- ---
--- ---
--- ---
+++ + + +
--- ---
--- ---
SEIZE Merebut
--- ---
--- ---
+++ + + +
+++ + + +
--- ---
--- ---
+++ + + +
+++ + + +
FORCE FORCE
2.441 2.441
+++ + + +
--- ---
--- ---
--- ---
--- ---
-2.487
+++ + + +
NONEVT
-2.083

c   c  
Schrodt: Forecasting Conflict
Page 62 Page 62
p *
#9pp$$"$.p$ ã$ã!$ã1$$
090 #.1.#$,('.#$ã ã$
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
('ã ã$ 
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
('ã ã$!$ã
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4

c   c  
Schrodt: Forecasting Conflict
Page 63 Page 63
p *È
#9pp$$"$.p$ ã$ã!$ã1$$
090 #.1.#$,'.#$ã ã$
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
'ã ã$ 
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
'ã ã$!$ã
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4

c  c 
Schrodt: Forecasting Conflict
Page 64 Page 64
p *)
#9pp$$"$.p$ ã$ã!$ã1$$090 #.1
.#$,('.#$ã ã$*
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
('ã ã$* 
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
('ã ã$*!$ã
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4

c 
c 

Schrodt: Forecasting Conflict


Page 65 Page 65
p *#9pp$$"$.p$ ã$ã!$ã1$$090 #
.1.#$,'.#$ã ã$*
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
'ã ã$* 
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4
BOS>
>BOS
CRO>
>CRO
SER>
>SER
KSV>
>KSV
'ã ã$*!$ã
-4--3
-3--2
-2--1
-1-0 -1-0
0-1 0-1
1-2 1-2
2-3 2-3
3-4 3-4

c 
Schrodt: Forecasting Conflict
Page 66
"
5   
  4
 ) 
 
T±tests were run for the difference between the 1-month and 6-month forecasting models.
There appear to be few if any systematic differences²in fact in comparisons of the later states,
the
number of significant t-values barely rises above the level expected by chance. This result is
Hasil ini
particularly strong in the low model estimates in .best models, which would suggest that the most
effective low models are simply modeling the background activity in non-conflict periods, with
appropriate adjustment to maximize the distinctions between this and the high-conflict periods.
In the high models, there is a consistent pattern is that there are more differences in the early
states of the models. This would suggest that more of the discrimination may be found in the
early
states, which is also consistent with how the HMM models themselves work: The best fit for a
sequence might involve simply cycling in the early states of the model, which should result in
more
precise estimation of those states. This is also consistent with the interpretation that these
forecasts
are dealing with general shifts in behavior (which involve relatively diffuse changes in the event
sequences) rather than a more finely-differentiated development of a six-stage crisis that would
involve a larger number of distinct Markov states.
Three Experiments in Simplifying the Model
As the discussion above makes clear, one major drawback of the HMM approach is the large
number of parameters. Because these parameters do not appear to exhibit any obvious
structure²widely different values produce roughly comparable results²it seems likely that
many
of them are redundant. I therefore undertook two experiments to test the effects on accuracy of
substantially reducing the number of parameters, first by looking only at the activity of the
dominant actor in the system²Serbia²and then by reducing the detail in the event codes. I also
Saya juga
looked at the effects of changing the classification variable from predicting conflict in a single
week to predicting conflict in a 1- or 2-month period.
Reducing the Number of Dyads: Serbia Only
Serbia is clearly the dominant actor in this crisis. During the 1990s, conflict in the former
Yugoslavia shifted focus from Croatia to Bosnia to Kosovo, but Serbia (and ethnic Serbs) were
involved at all of these stages. This suggested an alternative approach: just monitor the activity
of
Serbia. Serbia. Consequently I re-estimated a simpler model that involves only 45 codes:
[any source] -> Serbia
Serbia -> [any target]
plus the 00 nonevent code. The remainder of the forecasting design was the same as before.

c  c 
Schrodt: Forecasting Conflict
Page 67 Page 67
The focus on Serbia was also suggested by the fact that the model did not do very well at all on
predicting the outbreak of violence in Kosovo. This could be due to the fact that while the
original
model included events involving Kosovo, there was in fact no real precedent for violence in
Kosovo, and the reason that the model did not predict problems in Kosovo might have been due
to
it "looking" for the wrong thing.
Tables 3 and 4 show the summary statistics on the accuracy of the simplified model. In terms
Dalam hal
of overall accuracy, the results for the U and P models are almost identical to the results in the 4-
dyad model, and the N model actually improves by about 10%! (The improvement in the N
model
is due largely to increased accuracy during the implementation of the Dayton accords). These Ini
results are also mirrored in the analysis of the high and low-week errors in Table 4: The U and P
models gain between 5% and 20% observed accuracy in the high conflict week, while losing
only
about 2% observed accuracy in low-conflict weeks. The forecast accuracy of the U and P models
remains about the same. In the N model, the observed accuracy for high-conflict weeks increases
slightly (about 5% for the 3 and 6 month forecasts), and increases about 10% for low-conflict
weeks. minggu. Forecast accuracy increases about 7% for high-conflict weeks and stays the
same for low-
conflict weeks. In almost none of these cases is there a serious loss of accuracy when moving to
the simpler model.
18 18
Figure 7 shows the 5-week centered moving average of the 3-N and 3-P Serbia-only models;
in general these show patterns (including mirror-imaging) that are similar to those found in the 4-
dyad models. However, the direct comparison between the Serbia-only and 4-dyad 3-N
models²shown in Figure 8²reveals several interesting differences. First, as I had hoped, the
Serbia-only model is much quicker at picking up the cessation of hostilities following the Dayton
Accords²the two lines parallel each other during the period from July-1991 to October-1997,
with the Serbia-only model being about 30% more accurate. Second, there are two deep spikes of
low accuracy in the Serbia-only model that are not found in the 4-dyad model: these occur in late
1991 and late 1998. In both cases, the inaccuracy is due to false-positives: the model is saying
there will be conflict, but it is not reflected in the data. Both of these periods are followed by
major
outbreaks of violence.
This analysis suggests that this simple Serbia-only model may be more accurate than the more
complex 4-dyad model, despite the fact that the earlier model contains much more information.
There are at least three reasons that this might be true. First, simple models of social behavior are
often more accurate, because the measurement errors in an more complex model add more signal
18 18
The Serbia-only model also shows a greater likelihood of consistently being less accurate at
longer horizons,
which would be consistent with that model having reduced amounts of noise.

c 
c 

Schrodt: Forecasting Conflict
Page 68 Page 68
than noise.
19 19
(This parallels the experience of the State Failures Project, which started by looking
at several hundred variables, and found that all they really needed were half a dozen.) Second,
"watch the bad guys" is intuitively plausible²Serbia has been the initiator of much of the
violence
in this region, and when not the initiator, usually the target, so monitoring Serbia alone is going
to
get most of the required information. Such an approach might not work in an area characterized
by a truly multi-actor conflict such as central Africa, or the Afghan or Lebanese civil wars.
Finally, Akhirnya,
the underlying theory of sequence matching suggests that models should not be overly
specific²the whole point of the exercise is to generalize. In this case, the generalization is
across
all Serbian behavior, irrespective of target.
ã È ' ã È '


$
 :È . 
Best Models Model Terbaik
  
 
 
- 
-.È/")
- 
-.È
/")
1-U
88
3256 3256
80.1% 80,1%
32 32
13024
76.9% 76,9%
3-U
77
2786 2786
81.5% 81,5%
32 32
12736 12736
76.3% 76,3%
6-U
22
770 770
79.9% 79,9%
32 32
12320
74.4% 74,4%
1-P 1-P
77
2849 2849
81.1% 81,1%
32 32
13024
76.8% 76,8%
3-P 3-P
66
2388 2388
81.2% 81,2%
32 32
12736 12736
76.3% 76,3%
6-P 6-P
22
770 770
79.9% 79,9%
32 32
12320
75.0% 75,0%
1-N
31 31
12617 12617
64.1% 64,1%
32 32
13024
63.6% 63.6%
3-N
30 30
11940
59.7% 59,7%
32 32
12736 12736
58.9% 58,9%
6-N
32 32
12320
60.2% 60,2%
32 32
12320
60.2% 60,2%
19 19
In the case of event data, coverage bias is probably the most important source of error: reporters
go to where the
action is, and when the action is in Sarajevo, few reports will come out of Kosovo.

c 
 c 

Schrodt: Forecasting Conflict
Page 69 Page 69
ã È *
0 "  )1:È . 
!   ãÈ 
  
 
 
.È ã

p)  

.È ã

p)  

1-U
38.7%
51.5% 51,5%
43.1%
43.3% 43,3%
3-U
40.3% 40,3%
58.1% 58,1%
49.7% 49.7%
43.9% 43,9%
6-U
35.5% 35,5%
55.1% 55,1%
43.5% 43.5%
41.1% 41,1%
1-P 1-P
37.9% 37,9%
55.3% 55,3%
44.5% 44,5%
43.3% 43,3%
3-P 3-P
38.8% 38,8%
57.4% 57,4%
46.9% 46,9%
43.7%
6-P 6-P
33.1% 33,1%
55.6% 55,6%
42.8% 42,8%
42.1% 42,1%
1-N
92.8% 92,8%
35.5% 35,5%
93.0% 93,0%
35.2% 35,2%
3-N
91.3% 91,3%
33.1% 33,1%
91.7% 91.7%
32.7%
6-N
90.6% 90,6%
34.1% 34,1%
90.6% 90,6%
34.1% 34,1%
ã È *È
2"  )1:È . 
!   ãÈ 
  
 
 .È
p)  

.È ã

p)  

1-U
90.6% 90,6%
85.3% 85,3%
85.5% 85,5%
85.4% 85,4%
3-U
92.3% 92,3%
85.4% 85,4%
83.3% 83,3%
86.3% 86,3%
6-U
92.1% 92,1%
83.9% 83,9%
82.9% 82,9%
84.2% 84,2%
1-P 1-P
92.2% 92,2%
85.3% 85,3%
85.0% 85,0%
85.7% 85,7%
3-P 3-P
92.4% 92,4%
85.1% 85,1%
84.1% 84,1%
85.7% 85,7%
6-P 6-P
92.7% 92,7%
83.5% 83,5%
83.8% 83,8%
84.2% 84,2%
1-N
56.8% 56,8%
96.9% 96,9%
56.1% 56,1%
96.9% 96,9%
3-N
51.3% 51,3%
95.7% 95,7%
50.2% 50,2%
95.8% 95,8%
6-N
51.8% 51,8%
95.3% 95,3%
51.8% 51,8%
95.3% 95,3%

c 
 c 

Schrodt: Forecasting Conflict
Page 70 Page 70
p % 
È %
Accuracy of Serbia-only, 5-week Moving Average
00
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0.6 0.6
0.7 0.7
0.8 0.8
0.9 0.9
11
1991.01.01
1991.05.01
1991.09.01
1992.01.01
1992.05.01
1992.09.01
1993.01.01
1993.05.01
1993.09.01
1994.01.01
1994.05.01
1994.09.01
1995.01.01
1995.05.01
1995.09.01
1996.01.01
1996.05.01
1996.09.01
1997.01.01
1997.05.01
1997.09.01
1998.01.01
1998.05.01
1998.09.01
1999.01.01
3P 3P
3N 3N

c 
 c 

Schrodt: Forecasting Conflict
Page 71 Page 71
p & 
È &
Modifying the and Coding Scheme
A second experiment involves simplifying the coding system. Earlier work on conflict in the
Middle East (Gerner and Schrodt 1998) showed that in a cluster analysis, it was possible to
substantially reduce the number of coding categories substantially without much loss in
predictive
power (in fact predictive power might even be gained by eliminating sources of noise). I
therefore
re-estimated the models using the following five-category system:
0. 0. Non-event
1. 1. Verbal cooperation (WEIS categories 02, 03, 04, 05, 08, 09,10)
2. 2. Material cooperation (WEIS categories 01, 06, 07)
3. 3. Verbal conflict (WEIS categories 11, 12, 13, 14, 15, 16, 17)
4. 4. Material conflict (WEIS categories 18, 19, 20, 21, 22)
Comparison of 3N Models, 5-week Moving Average
00
0.1 0.1
0.2 0.2
0.3 0.3
0.4 0.4
0.5 0.5
0.6 0.6
0.7 0.7
0.8 0.8
0.9 0.9
11
1991.01.01
1991.05.01
1991.09.01
1992.01.01
1992.05.01
1992.09.01
1993.01.01
1993.05.01
1993.09.01
1994.01.01
1994.05.01
1994.09.01
1995.01.01
1995.05.01
1995.09.01
1996.01.01
1996.05.01
1996.09.01
1997.01.01
1997.05.01
1997.09.01
1998.01.01
1998.05.01
1998.09.01
1999.01.01
3N-Serbia
3N-4 Dyads

c 
 c 

Schrodt: Forecasting Conflict
Page 72
This reduces the total number of codes in the 8-dyad Balkans model from 177 to 33, which also
substantially reduces the number of low frequency code categories. (This is also likely to reduce
the effect of coding variance and coding error somewhat: Several of the ³verbal conflict´ codes
in
WEIS are ambiguous even for human coders, and the automated coding probably generates some
misclassification in those categories.)
The results of this experiment are given in Table 13a. As before, the models were evaluated
with 1, 3 and 6-month forecast lags, and with the P and N weighting scheme. 32 Monte-Carlo
genetic algorithm estimates were done for each set of experimental parameters. For purposes of
Untuk tujuan
comparison, Table 13b presents the statistics for the original model in the same format.
In general, the results of the new analysis are comparable to those of the original analysis. As
Seperti
before, the drop-off in accuracy with the increasing forecasting lag is small²about 4% from the
1-
month to 6-month forecast lag²but consistently there is a small decrease. The overall accuracy
measure decreases about 4% for the P models and  m   about 8% for the N models. The The
largest difference in the results occurs with respect to the accuracy of the high-conflict
predictions
in the P models²these average about 18% better in the percentage of the observed high weeks
that
were correctly forecast, albeit at the cost of an 8% decrease in the corresponding percentage of
the
observed low weeks that were correctly forecast. The N model shows an 11% increase in the
percentage of the observed low weeks that were correctly forecast and a 5% increase in the
percentage of forecasts of high conflict that actually had high conflict. All of the remaining
statistics differ from the original model by less than 3%.
This analysis clearly supports the results found in Gerner and Schrodt (1998)²the use of
simplified event coding systems at worst involves only a small penalty in terms of predictive
accuracy, and at best can actually improve the accuracy, probably through the reduction of noise.
This is particularly important when automated coding is being used, since automated coding is
generally less capable of making subtle distinctions between event categories, but generally is
quite
good at making large distinctions such as the difference between cooperative and conflictual
behavior. perilaku.
20 20
20 20
It should be noted that both this test and the earlier Gerner and Schrodt (1998) test use machine-
coded data, so
this effect might be due to the errors found in that type of coding. However, this seems
somewhat unlikely
given the magnitude of the effect, the fact that the overall error rate in machine coding is
comparable to that of
human coding, and the fact that many of the categories that are ambigous in machine coding are
also ambiguous
to human coders.

c 
 c 

Schrodt: Forecasting Conflict
Page 73 Page 73
ã È '  )) )3" " 

Experiment Percobaan
%accuracy
%high
correct benar
%low
correct benar
%high
forecast ramalan
%low
forecast ramalan
P1 P1
74.4 74.4
46.2 46.2
81.5 81.5
38.9 38.9
85.6 85.6
P3 P3
71.7 71.7
44.1 44.1
78.9 78.9
35.4 35.4
84.4 84.4
P6 P6
71.4 71.4
44.2 44.2
78.8 78.8
36.4 36.4
83.8 83.8
N1 N1
61.9 61.9
90.7 90.7
54.6 54.6
33.7 33.7
95.8 95.8
N3 N3
57.8 57.8
87.0 87.0
50.2 50.2
31.4 31.4
93.6 93.6
N6 N6
56.8 56.8
85.9 85.9
48.8 48.8
31.5 31.5
92.7 92.7
ã È 'È )) )1( )  (  '" " 

 

Experiment Percobaan
%accuracy
%high
correct benar
%low
correct benar
%high
forecast ramalan
%low
forecast ramalan
P1 P1
77.6 77.6
29.3 29.3
89.5 89.5
40.8 40.8
83.7 83.7
P3 P3
76.0 76.0
29.0 29.0
87.9 87.9
37.9 37.9
82.9 82.9
P6 P6
76.9 76.9
25.9 25.9
90.6 90.6
42.6 42.6
82.0 82.0
N1 N1
54.2 54.2
92.7 92.7
45.3 45.3
28.1 28.1
96.4 96.4
N3 N3
49.0 49.0
88.1 88.1
39.6 39.6
25.9 25.9
93.3 93.3
N6 N6
47.7 47.7
88.5
37.4 37.4
26.3 26.3
92.8 92.8
Modifying the Prediction Framework
The final experiment involves modifying the criteria for making a prediction. As noted at a
number of points above, the objective of predicting the level of conflict in a single week is
unrealistically precise²most predictions by human political analysts are made over more
general
periods of time. periode waktu. Consequently, the classification variable was modified to
indicate whether a high-
conflict week occurred during a 4- or 8-week period following the forecasting-lag period. The
The
threshold for high-conflict was set at greater than 26 WEIS category  events for the 4-week
c 
c 

Schrodt: Forecasting Conflict
Page 74 Page 74
period and greater than 57 WEIS category  events for the 8-week period; approximately 30%
of
the periods fall into the high-conflict category.
These results are reported in Table 14, the 4 and 8 prefixes in the ³Experiment´ code refer to
conflict occurring in 4- and 8-week periods respectively. These runs were done with the 5-code,
rather than the earlier 22-code, system; and there were 32 Monte Carlo experiments for each
experimental condition.
Surprisingly, the overall accuracy changes very little from the results of the 5-code experiment
reported in Table 13a; almost all of the accuracy percentages are within 3% of the earlier results.
However, there are substantial changes in the   of the accuracy. As with the change in the
5-
code system, this change benefits the accuracy of the forecasts of the high-conflict months. In the
Dalam
P model, the number of high conflict periods correctly predicted increases by about 13% for the
1-
month and 3-month forecast lags, and the percentage of high conflict periods where high conflict
is
actually observed also increases by about 13%. In the N model, the percentage of high conflict
periods where high conflict is actually observed also increases by about 10%, but there is almost
no change in the the number of high conflict periods correctly predicted. Because the overall
accuracy changes little, these increases in accuracy for the high-conflict periods are compensated
by decreases in accuracy for the low-conflict periods, though those percentage changes are
smaller²typically around 4%²because the number of low-conflict periods is larger.
These results were somewhat unexpected: I had expected that predictions for the longer
periods would be more accurate because they would not be subject to errors due to the random
week-to-week variations within a period of high conflict. In a sense, this is what occurs in the
higher prediction accuracy for high-conflict periods. However, that has relatively little impact on
the overall accuracy, which is dominated by the low-conflict periods and low-conflict is highly
autocorrelated. The results are consistent with the model is picking up very general
characteristics
of the behavior that change only slowly over time, rather than looking at indicators for specific
weeks, despite that choice of the classification variable; this will be discussed in greater detail
below. di bawah ini.

c 

c 

Schrodt: Forecasting Conflict


Page 75 Page 75
ã È * ã È * )) )* &1( )  (  3" " 

 

Experiment Percobaan
%accuracy
%high
correct benar
%low
correct benar
%high
forecast ramalan
%low
forecast ramalan
4P1
70.7 70.7
59.6 59.6
75.5 75.5
51.2 51.2
81.3 81.3
4P3
69.1 69.1
55.0 55.0
75.4 75.4
49.9 49.9
79.1 79.1
4P6
67.1 67.1
47.3 47.3
75.6
45.4 45.4
77.1 77.1
8P1
72.8 72.8
59.4 59.4
78.2 78.2
52.3 52.3
82.8 82.8
8P3 8P3
69.6 69.6
58.9
74.0 74.0
48.4 48.4
81.3 81.3
8P6
69.7 69.7
51.6 51.6
77.5 77.5
49.2 49.2
79.1 79.1
4N1
62.1 62.1
89.2 89.2
50.4 50.4
43.7 43.7
91.6 91.6
4N3
56.8 56.8
87.4 87.4
43.2 43.2
40.6 40.6
88.6 88.6
4N6
55.3 55.3
85.4 85.4
42.5 42.5
38.9 38.9
87.2 87.2
8N1
64.9 64.9
89.6 89.6
55.0 55.0
44.5 44.5
93.0 93.0
8N3
59.1 59.1
86.9 86.9
47.6 47.6
40.7 40.7
89.8 89.8
8N6
58.5 58.5
86.3 86.3
46.8 46.8
40.7 40.7
89.0 89.0
Conclusion Kesimpulan
The overall conclusion of this analysis is that hidden Markov models are a robust, though
hardly flawless, method for forecasting political conflict, at least when applied to area such as
the
former Yugoslavia where substantial information about political events is available. From the
Dari
standpoint of pure prediction, the models are credible.
Unfortunately, these models are less useful from the standpoint of    m ; in particular, it is
very difficult to figure out what information the HMM is using to make these prediction. At Pada
various points in the discussion above, I have noted some clear²and generally plausible²
patterns
in the estimated probabilities. However, in general it has proven quite difficult to make much
sense
out of these. This is not to say that this exercise has been useless²without looking for patterns,
there was no way to know they were not present²but a series of experiments trying to find
patterns in these coefficients has not produced any obvious results.

c 
c 

Schrodt: Forecasting Conflict
Page 76 Page 76
There are a couple of likely reasons for this. First, these forecasting models have relatively
long time horizons. The common-sense signs of short-term escalatory behavior such as demands,
threats and small-scale incidents of violence will not necessarily be found at three and six month
horizons or, in this data, even at one month. (One sees a little bit of this, for example in
behaviors
coming from Serbia, or towards Bosnia and Kosovo, but not a lot).
What the models seem to be picking up instead are relatively diffuse indicators, and often as
not, simply increased attention to the area by the international media (as well as the international
community. masyarakat. For example, one would expect that if a NATO commander or UN
representative
called attention to some issue, Reuters would almost always report it, whereas if the mayor of a
village made the same comment, this might be ignored. Some of these indicators may be
indirect:
For example I initially viewed the emphasis in the high-conflict models on Kosovo to be a
statistical artifact, but this could also be reflecting that fact that after 1991 or so, the international
community consistently responded to aggressive moves by Serbia by warning Serbia not to do
anything in Kosovo, and these activities are probably picked up in the data set. The presence or
Keberadaan atau
absence of non-events is clearly very important, a result that was expected²the fact that the
international media are reporting on an area is by itself a useful indicator.
The interpretation of the coefficients is further complicated by the fact that we are dealing with
probabilities²which by definition sum to 1.0²so an increase in one probability necessarily
leads
to a decrease in another. Add to this the propensity of interactions in event data to be
symmetrical,
and the fact that the codes are aggregated not only within the four actors in the Balkans, but also
interactions between those actors and the international system as a whole, and the determination
of
the parameters gets very complex. In particular, it is quite different from the more familiar task
of
interpreting regression coefficients, which (in the absence of significant collinearity) are more or
less independent of each other in value.
This is further complicated by the problem of the indeterminancy of the estimates produced by
the Baum-Welch algorithm. This indeterminancy does not seem to be dealt with in detail in the
HMM literature²most HMM applications are solely concerned with prediction, not
inference²but where it is mentioned, the experience that I have had estimating these models
appears to be typical. I have done a series of additional unreported experiments to attempt to find
ways to reduce the variance of the estimates²increasing the number of templates used, varying
the
parameters of the genetic algorithm, changing the convergence conditions of the Baum-Welch
algorithm, and setting the initial observation probabilities in the vicinity of the marginal
probabilities of events in the data set as a whole²and none of these had a major impact. There is
Ada
some limited evidence that the variance in the mm m of the models estimated with the 5-code
system is less than that of the 22-code system, but the variance in the parameters is still quite
high.

c 
c 

Schrodt: Forecasting Conflict
Page 77 Page 77
Finally, the sheer number of parameter estimates generated by these HMMs complicates the
problem of interpretation. The HMM models may be similar to neural networks in this regard:
the
diffuse coefficient structure is the model's way of dealing with the high degree of uncertainty in
the
underlying data, and the complexity of tradeoffs between the parameter values makes them
almost
impossible to interpret in a simple fashion. Doing a full interpretation would ideally involve five
dimensions of comparison²WEIS category, dyad, Markov state, weighting scheme and
high/low
model²which is three dimensions more than most people can deal with. I've focused here on
two
weights, one or two Markov states, and a two-dimensional actor-by-code comparison, but that
obviously leaves plenty of other possibilities that have not been explored.

c  c 
Schrodt: Forecasting Conflict
Page 78 Page 78
References Referensi
Allison, GT 1971. ã  m   m  Boston: Little, Brown.
Anderson, PW, KJ Arrow and D. Pines, eds. 1988. ã m    !  
"  New York: Addison Wesley.
Azar, EE, and T. Sloan. 1975.    #  m . Pittsburgh: University Center for
International Studies, University of Pittsburgh.
Bartholomew, DJ 1971. " mm$  " m m   New York: Wiley.
Bennett, S. and PA Schrodt. 1987. 1987. Linear Event Patterns in WEIS Data. Paper presented at
Makalah disampaikan pada
American Political Science Association, Chicago.
Bloomfield, LP, and A. Moulton. 1989. !%"!&'###!   % "  %  
( m! m Cambridge: MIT Center for International Studies.
Bloomfield, LP and A. Moulton. 1997. $  #    ! m New York: St.
Martin's Press. Martin's Press.
Bueno de Mesquita, B., D. Newman and A. Rabushka. 1996. ) *  + , 
Chatham, NJ: Chatham House Publishers.
Bueno de Mesquita, B. 1981. ã -ã New Haven: Yale University Press.
Butterworth, RL 1976. $  #   ! m
./0120 "   Pittsburgh:
University of Pittsburgh University Center for International Studies.
Casti, JL 1997. -  3 -   New York: Wiley.
Choucri, N. and TW Robinson, eds. 1979. *  m  #    )  ã 

$  
  
  m San Francisco: WH Freeman.
Cimbala, S. 1987. %m#   m  ' " m  Lexington, MA: Lexington
Books.
Cyert, RM and JG March. 1963. %3  ã   * . Englewood Cliffs, NJ:
Englewood Cliffs, NJ:
Prentice-Hall. Prentice-Hall.
Feller, William. 1968. % #  m   ã  #%m  New York:
Wiley. Wiley.
Gerner, DJ, PA Schrodt, RA Francisco, and JL Weddle. 1994. 1994. The Machine Coding of
Events from Regional and International Sources. #    "  4   38:91-
119. 119.
Gerner, DJ and PA Schrodt. 1998. 1998. ³The Effects of Media Coverage on Crisis Assessment
and Early Warning in the Middle East.´ In -   )   , ed.
Susanne Schmeidl and Howard Adelman. New York: Columbia University Press-
Columbia International Affairs Online.
Gurr, TR and B. Harff. 1996. -   !  ! m +   !
Tokyo: United Nations University Press, Monograph Series on Governance and Conflict
Resolution. Resolusi.
Goldstein, JS 1992. A Conflict-Cooperation Scale for WEIS Events Data. 5   ! m
)    36: 369-385.
Hopple, GW, SJ Andriole, and A. Freedy, eds. 1984. ' " m !*  m 
 $    .  $    Boulder: Westview.

c   c  
Schrodt: Forecasting Conflict
Page 79 Page 79
Hudson, V., ed. 1991. %m#   m  #     m . Boulder: Westview
Hughes, BB 1984. -  *   %!m%  %    Baltimore: Johns
Hopkins. Hopkins.
Huxtable, PA 1997. "Uncertainty and Foreign Policy-making: Conflict and Cooperation in
West Africa." Ph.D. dissertation, University of Kansas.
Huxtable, PA and JC Pevehouse. 1996. 1996. Potential Validity Problems in Events Data
Collection. #    "  '   21: 8-19.
Kauffman, SA 1993. ã &  &  Oxford: Oxford University Press.
Khong, YF 1992. %   - . Princeton: Princeton University Press. Princeton: Princeton
University Press.
Kruskal, JB 1983. An Overview of Sequence Comparison. In ã -
"  
$m   m  
ed. D. Sankoff and JB Kruskal. New York: Addison-Wesley. New York:
Addison-Wesley.
Laurance, EJ 1990. "Events Data and Policy Analysis."  m"m m  23:111-132.
Lebow, RN 1981. 3   m  -ã '   #    !  Baltimore:
Johns Hopkins. Johns Hopkins.
Leng, RJ 1987. 3  !    -
.6.7./21 . (ICPSR 8606). Ann Arbor: Inter-
university Consortium for Political and Social Research.
Leng, RJ 1993. #   !3  
.6.7./68 New York: Cambridge University Press.
Lund, MS 1996.    9  ! m%"       m
Washington, DC: United States Institute for Peace.
McClelland, CA 1976. -   :#  m "  !   (ICPSR 5211). Ann Arbor:
Ann Arbor:
Inter-University Consortium for Political and Social Research.
May, E. 1973. ;(  ;  ã <  $   +  % m *  
 m New York: Oxford University Press.
Mefford, D. 1985. Formulating Foreign Policy on the Basis of Historical Programming. In
Dalam
 m$  #    ! m , ed. U. Luterbacher and MD Ward  Boulder:
Lynne Rienner Publishing.
Merritt, RL, RG Muncaster, and DA Zinnes. 1993. #      
    #) ## . Ann Arbor: University of Michigan Press. Ann Arbor:
University of Michigan Press.
Myers, R. and J. Whitson. 1995. 1995. HIDDEN MARKOV MODEL for automatic speech
recognition (C++ source code).
http://www.itl.atr.co.jp/comp.speech/Section6/Recognition/myers.hmm.html
Neustadt, RE and ER May. 1986. ã   ã ã <  +   m
$  New York: Free Press.
Rabiner, LR 1989. A Tutorial on Hidden Markov Models and Selected Applications in Speech
Recognition.  m    # 77,2:257-286
Sankoff, D. and JB Kruskal, eds. 1983. ã -
"   $m   m  ã
ã  mm  "  m !   New York: Addison-Wesley.
Schrodt, PA 1985. The Role of Stochastic Models in International Relations Research. In Dalam
ã  
$  "   #    )   , ed. MD Ward. Boulder: Boulder:
Rienner

c   c  
Schrodt: Forecasting Conflict
Page 80 Page 80
Schrodt, PA 1990. Parallel Event Sequences in International Crises, 1835-1940.  m
3   12: 97-123.
Schrodt, PA 1991. Pattern Recognition in International Event Sequences: A Machine Learning
Approach. In %m#   m  #     m , ed. V. Hudson. Boulder: Boulder:
Westview. Westview.
Schrodt, PA 1993. Rules and Co-Adaptation in Foreign Policy Behavior. Paper presented at
Makalah disampaikan pada
the International Studies Association, Acapulco.
Schrodt, PA 1994. Event Data in Foreign Policy Analysis. in L. Neack, JAK Hey, and PJ
Haney. *    m% !   !   New York: Prentice-Hall, pp.
145-166. 145-166.
Schrodt, PA 1999. ³Early Warning of Conflict in Southern Lebanon using Hidden Markov
Models.´ In ã <    $    = 9  m , ed. Harvey Starr.
Pp. Pp. 131-162. 131-162. New York: St. Martin's Press, 1999.
Schrodt, PA 2000. ³Pattern Recognition of International Crises using Hidden Markov
Models.´ In  m!  '  $   m , ed. Diana Richards. Pp. Pp.
296-328. Ann Arbor: University of Michigan Press. Ann Arbor: University of Michigan Press.
Schrodt, PA and DJ Gerner. 1994 . 1994. Validity assessment of a machine-coded event data set
for
the Middle East, 1982-1992. % m 5    m"m m 38: 825-854.
Schrodt, PA, SG Davis and JL Weddle. 1994. 1994. Political Science: KEDS²A Program for
the Machine Coding of Event Data. " m"m m !   )   12: 561-588.
Schrodt, PA, and DJ Gerner. 1997. 1997. Empirical Indicators of Crisis Phase in the Middle
East,
1982-1995. 5   ! m)    41:529-552.
Schrodt, PA, and DJ Gerner. 1997. 1997. An Event Data Set for the Arabian/Persian Gulf Region
1979-1997. Paper presented at the International Studies Association, Minneapolis, March
1998.(this paper be downloaded from http://www.ukans.edu/keds/papers.html)
Sherman, FL, and L. Neack. 1993. 1993. Imagining the Possibilities: The Prospects of Isolating
the
Genome of International Conflict from the SHERFACS Dataset. In #     
     #) ## . ed. RL Merritt, RG Muncaster, and DA
Zinnes. Ann Arbor: University of Michigan Press. Ann Arbor: University of Michigan Press.
Singer, JD and Wallace MD 1979. ã %  - -  # m  - 
 m .   Beverly Hills: Sage. Beverly Hills: Sage.
Van Creveld, M. 1991. ã m   - . New York: Free Press. New York: Free Press.
Vertzberger, YI 1990. ã -    $ #    m  
!   
 m   *    m m $  Stanford: Stanford University Press.