Sei sulla pagina 1di 35

Robert Engle and Robert Ferstenberg

Microstructure in Paris
December 8, 2014

Is varying over time and over assets


Is a powerful input to many financial
decisions such as portfolio construction and
trading algorithms
Is never directly observed but is believed to
be correlated with volatility and volume and
bid ask spread
Can be interpreted as the ease with which an
investor can purchase or sell large quantities.
Forecasting Transaction Costs

Create several measures of transaction costs


based on market data with daily market
volume orders (DMVO)
Produce forecasts of these measures
Compare these forecasts with transaction
records from Abel Noser Solutions (ANS)
Do this daily for many US equity names over a
15 year period.

IS is the difference between the market price


at the time the order is entered and the price
at which a trade is executed, measured as a
return.
IS =log(execution price/arrival price)*side
Where side is 1 for a buy order and -1 for a sell
order
IS should be positive on average for all trades.
IS has low ratio of signal to noise.

Perold, Kyle, Grinold and Kahn, Engle


Ferstenberg Russell, Kyle and Obisheva,
Pastore and Stambaugh, Hasbrouck, Roll,
Amihud, Easley Lopez de Prado and OHara,
Russell, Jones and Lipson, Engle and Lange,
Keim and Madhavan.
Many industry studies, ITG, ANS, AQR
Apologies for the authors omitted

Forecast Transaction Costs

Use transaction data from ANS. These are


parent orders which were typically filled with
sequences of small child orders.
We have execution price and fill quantity. We
also have a buy sell indicator.
We approximate the arrival price by the opening
price.
We do not have the order size for orders that are
not completed. This is probably an important
bias.
Thus we can compute IS for more than 300
million executed orders since 1998.

We consider the entire daily market volume on a


name as an order.
The arrival price is then the open and the
execution price is VWAP. We approximate with
the close.
The direction of the trade must be inferred from
the price movement. If the price goes up, the
trade is classified as a buy and conversely for a
sell.
An order which is a small fraction of daily volume
will incur only a fraction of the daily impact.

Trade direction has always been inferred from


prices with market data. After all, every
execution has a willing buyer and a seller.
The Lee and Ready algorithm offers two solutions
to signing a single trade either the price change
is used or the price relative to the recent midquote.
Lee and Ready was motivated by the idea that a
market order was active and executed by a
specialist. Today there are no specialists and
informed traders do not generally use market
orders. Limit orders can have price impact so the
direction of trade can best be inferred by the
price movement.

Not all trades have equal price impact. More


informative trades have more impact. Hence,
it makes sense to measure impact on
average. Information on the informativeness
of trades can be used if available.
Easley, Lopez de Prado, OHara in a series of
papers have introduced Bulk Volume
Classification or BVC which uses inter daily
data to classify trades by the price change.
We take this to the daily limit. DMVO

Volatility
Volume
Average Daily Volume
Bid Ask Spread
The correlation of volume and volatility is
central to price impact.
Predictions of transaction cost are based on
order size relative to predicted volume.

Amihud, Yakov(2002), Illiquidity and stock


returns: Cross-section and time series effects,

Journal of Financial Markets 5, 3156.

Grinold,Richard and Ronald Kahn(1999) Active

Portfolio Management: A Quantitative Approach


for Producing Superior Returns and Controlling
Risk, McGraw Hill
Engle, Robert, Robert Ferstenberg and Jeffrey
Russell (2012), Measuring and Modeling
Execution Cost and Risk, The Journal of Portfolio

Management.

Historical forecasts
Lagged 21 day average of absolute return divided
by daily dollar volume
There are no unknown parameters in this system
Forecasting can also be done econometrically by
using MEM models as in VLAB

Transaction costs are approximated by


multiplying ILLIQ by the order in dollars

Volumet
ISt = b1Spreadt- 1 + b 2Volatilityt- 1 *
ADVt- 1

Where Volatility is 21 day lagged standard


deviation of open to close returns, Volume is
todays dollar volume, ADV is 21 day average
dollar volume and Spread is the bid ask
spread divided by open price.
Estimated with panel of CRSP names for a
year and forecast with the same betas for the
following year.

b0

b1
t- 1

ISt = e Volatility

b2

b3
t- 1

Volumet ADV

This specification ensures that expected IS is


non-negative. It is a generalization of GK.
Estimate with panel of one year data on all
trades in the year.
Forecast for the next years trades.

Mean by Year of Daily ADN Weighted Mean of Mean of 21 Day Lagged IILIQ
2,50E-08

2,00E-08

1,50E-08

1,00E-08

5,00E-09

0,00E+00
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013

Mean by Year of Daily ADN

Mean by Year of Daily ADN

Weighted Mean of Out of Sample

Weighted Mean of Out of Sample

ILLIQ Forecast of 1% ADV

ILLIQ Forecast of $1M

0,0005

5,00E-06

0,00045

4,50E-06

0,0004

4,00E-06

0,00035

3,50E-06

0,0003
0,00025
0,0002
0,00015
0,0001
0,00005
0

3,00E-06

2,50E-06
2,00E-06
1,50E-06
1,00E-06
5,00E-07
0,00E+00
1

NYU Stern Volatility Institute

11

13

15

12/22/2014

17

19

21

16

0,035

0,03

0,025

0,02

forecast
realization

0,015

prediction

0,01

0,005

0
1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013

NYU Stern Volatility Institute

12/22/2014

18

Mean by Year of Daily ADN

Mean by Year of Daily ADN

Weighted Mean of Out of Sample

Weighted Mean of Out of Sample

CRSP GK Forecast of 1% ADV

CRSP GK Forecast of $1M

0,004

0,0025

0,0035
0,002

0,003
0,0025

0,0015

0,002
0,001

0,0015
0,001

0,0005

0,0005
0

0
1994 1996 1998 2000 2002 2004 2006 2008 2010 2012

1994 1996 1998 2000 2002 2004 2006 2008 2010 2012

NYU Stern Volatility Institute

12/22/2014

19

0,035

0,03

0,025

0,02

forecast
realization

0,015

prediction

0,01

0,005

0
1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013

NYU Stern Volatility Institute

12/22/2014

21

Mean by Year of Daily ADN

Mean by Year of Daily ADN

Weighted Mean of Out of Sample

Weighted Mean of Out of Sample

CRSP EFR Forecast of 1% ADV

CRSP EFR Forecast of $1M

0,008

0,0014

0,007

0,0012

0,006

0,001

0,005
0,0008

0,004
0,0006
0,003
0,0004

0,002

0,0002

0,001
0

0
1994 1996 1998 2000 2002 2004 2006 2008 2010 2012

1994 1996 1998 2000 2002 2004 2006 2008 2010 2012

NYU Stern Volatility Institute

12/22/2014

22

ABEL NOSER SOLUTIONS


ANCERNO TRANSACTION
DATA SET

More than 300 million orders executed by


ANS clients from 1998-2013.
Arrival price is the open
Execution price is the average price of the
executed shares.
Fill is the number of shares executed.
Data are cleaned, extremes are removed and
bucketed in many ways.
Data are matched with CRSP by
date/cusip/symbol

Mean of IS 50-51%tile

Mean IS
0,003

0,0045
0,004

0,0025

0,0035
0,002

0,003

0,0025

EW-IS

0,002

VW-IS

0,0015

AW-IS

0,001

EW-IS

0,0015

VW-IS
AW-IS

0,001

NW-IS

NW-IS
0,0005

0,0005
0

-0,0005
-0,0005
-0,001

NYU Stern Volatility Institute

12/22/2014

26

ANS Summary Statistics By Year


Comparison between:
VWAP: Un-weighted
average execution cost if
ANS fills are priced at
VWAP approximated as
average of open and close
price
EW-IS: Un-weighted
average reported by ANS

0,0035

0,003

0,0025

0,002
EW-IS
VWAP

0,0015

0,001

0,0005

0
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013

Assumption is that
impact of ANS executions
are priced in the market
data

Mean IS Fills Priced at VWAP

12/22/2014

NYU Stern Volatility Institute

29

The ANS estimate of Implementation Shortfall


shows some surprising features.
The peak IS for medium to large trades
occurs in 2009 and then it falls slowly.
For smaller trades it is in 2011 and seems
surprisingly large.
2001 and 2002 are very low cost years and
this seems surprising too.

For each trade in ANS data set, predict IS


from model parameters in preceding year
using ANS fill and market variables.
Predict ANS IS as an intercept plus slope
coefficient on forecast for all trades in a year.
This is sometimes called Mincer-Zarnowitz
regression.
Consider the predictions of this model as the
transaction cost model.

0,0025

0,002

0,0015
forecast
realization
prediction

0,001

0,0005

0
1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

NYU Stern Volatility Institute

2013

12/22/2014

32

0,0035

0,003

0,0025

0,002

forecast
realization

0,0015

prediction

0,001

0,0005

0
1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

NYU Stern Volatility Institute

2013

12/22/2014

33

Better spread data


Explore impact of estimation universe
CRSP is ~5000 names
ANS is ~2000 names

Explore models of excess returns


Seek alternative samples of executions.

A PROMISING APPROACH
TO LIQUIDITY FORECASTING!

Potrebbero piacerti anche