APPLICATION OF MACHINE LEARNING

IN DEMAND FORECASTING

AE 407: PRODUCTION AND OPERATIONS

MANAGEMENT
SUBMITTED BY:
ABBIMNYOU WARIKOO (2K16/AE/001)
ARNAB BISWAS (2K16/AE/017)
DHRUV SHARMA (2K16/AE/029)
KRISHNA DASH (2K16/AE/038)
 CONTENTS

 Abstract.

 Introduction.

 Case study:

 A Review on AI- Based Demand Forecasting for Both Reliable Forecasting and
Efficient Operation: Dynamic Ensemble Forecasting.
 A Review on Machine Learning-Based Demand Forecasting in Supply Chains.
 A Review on Machine Learning Methodology Applied to Demand and Sales
Forecasting.
 A Review on Application of Machine Learning techniques for Supply Chain
Demand Forecasting.

 Conclusion.

 Future Scope of Machine Learning in Demand Forecasting.

 ABSTRACT

Demand forecasting is one of the main issues of supply chains. It aimed to optimize stocks, reduce costs, and
increase sales, profit, and customer loyalty. For this purpose, historical data can be analyzed to improve demand
forecasting by using various methods like machine learning techniques, time series analysis, and deep learning
models. In this report, we have prepared literature reviews on various research papers proposed on the application
of machine learning in supply chain demand forecasting and its improved machine learning models for reliable
forecasting and efficient operation of an organisation. These improved models are based on the analysis and
interpretation of the historical data by using different forecasting methods which includes time series analysis
techniques, regression algorithms, and deep learning models. This paper also introduces the dynamic ensemble
forecasting technology and attributes decomposition model technology developed by Fujitsu laboratories to provide
solutions to the challenge of demand forecasting in the distribution industry. It also describes solutions based on
these technologies and actual application cases. As demand forecasting is statistics-heavy and data-rich, it is ripe for
the application of machine learning algorithms. The automation of data flows helps to manage logistics and find new
ways to optimize an organization’s performance.
 INTRODUCTION
Demand forecasting is a field of predictive analytics that aims to predict the demand of customers. It is done by
analyzing statistical data and looking for patterns and correlations. Machine learning takes the practice to a higher
level. According to McKinsey study, AI-enhanced supply chain management may lead to improved accuracy by
reducing forecasting errors by 20-50%.
Why demand forecasting is essential? A company faces some challenges when the real demand of a product varies:
 Income and profit loss when a product is out of stock or a service is unavailable
 Cash tied up in stock
 The reduced margins that come with getting it out of warehouses.
Since competition is increased day by day among retailers at the market, companies are focusing more predictive
analytics techniques in order to decrease their costs and increase their productivity and profit. Considering
competition and financial constraints in distribution industry, it is very crucial to have an accurate demand
forecasting and inventory control system for management of effective operations.
Unlike in the past, when anything that was produced would sell, it is becoming increasingly difficult to accurately
forecast demand because markets are now saturated and consumers can choose from so many choices. Demand is
predicted using both qualitative and quantitative methods. In the most basic form, it is done by leveraging the
experience of the seller or collective brainstorming in the company. Conducting market surveys is another method of
demand forecasting and applied in some major industries. As companies are relying more and more-on the data they
collect, it is no surprise that traditional methods are being replaced by data-powered, automated systems.
Traditional forecasting methods are based on time series. This means that they are applied under the hypothesis
that the past demand can statistically estimate the future demand. Using big data demand prediction is enabling a
wide range of companies to leverage machine learning models in data exploration. As AI technologies evolve, their
use is increasingly being looked to as a solution for demand forecasting.
“You can reduce your planned inventory by 2.5%, if you improve your inventory forecast by 1%”
In this report, we have reviewed some research papers which proposed various machine learning models to improve
demand forecasting process. The various machine learning models are based on the analysis of past data which uses
time series analysis and vector regression algorithms to optimize the demand forecasts. It also describes dynamic
ensemble forecasting technology for products which are already in the market and attribute decomposition model
technology for new products which are about to make an introduction in the market. Experiments are conducted to
evaluate the accuracy of demand forecasting for products using past data from manufacturers, retailers and
wholesalers.
Accurate forecasting of the demand of the products reduces forecasting errors and inventory cost due to planned
demand and inventory forecasting. It helps an organisation to meet the demand of the product in the market and
reducing dead stock. This, in turn results in increased sales and productivity of a company. As a result, accurate
demand forecasting in an organisation causes improvement in sales and business success.
 A Review on AI-Based Demand Forecasting for Both Reliable
Forecasting and Efficient Operation: Dynamic Ensemble Forecasting
 Krishna Dash (2K16/AE/038)

INTRODUCTION
In the distribution industry, where manufacturers, wholesalers and retailers make up the supply chain, the
forecasting of demand majorly affects the business success of an organisation. Business activities such as
product planning, product management, logistics and sales are also based of the forecasting of demand.
Therefore, it is essential for companies to forecast the demand for products more accurately for business
success. Presently, companies use all kinds of information and data to forecast demand.

In recent years, it has become extremely difficult for companies to forecast demand accurately because of the
increase in the market competition for a product. Nowadays, consumers can choose from a variety of products
available in the market, which makes it difficult to forecast demand. So, some ways must be found to predict this
fluctuation in demand of a product.
As AI technologies evolve, their use is increasingly being looked to as a solution for demand forecasting. Fujitsu
laboratories have developed the Dynamic Ensemble forecasting technology and Attribute Decomposition Model
technology to provide solutions to the challenge of demand forecasting in the distribution industry.

LITERATURE REVIEW
 Current State of Demand Forecasting:
Some companies calculate forecast values manually using spreadsheet software, others use statistical
methods and demand forecasting systems by machine learning for planning work. Various methods of
demand forecasting have been in use recently. Table 1 lists representative forecast methods and
characteristics.

Forecast accuracy Predictor

Magnitude of number to be When Possibility of Ease of
Forecast method Description forecasted volume of incorporation understandi
Large Small learning into the ng of
data is model influence of
small factors and
its extent
Random walk The simplest method which use the Low High High No -
latest actual value as the prediction
value.
Multiple Regression Prediction from correlation between Medium Low Low Yes Easy
factors (holidays, weather, etc.) and
sales figures.
Poisson Regression This method allows accurate prediction Medium High Low Yes Easy
for products with low actual value
volume.
ARIMA Prediction from correlation with past Medium Low Low No Difficult
actual values.
ARIMAX A combination of the ARIMA model and High Low Low Yes Easy
multiple regressions.
Dynamic linear Each time data is added, the internal High Medium High Yes Easy
model latent state is updated and prediction is
executed accordingly. This model can
handle dynamic change.
Exponential Prediction focusing on trends and Medium Low High No -
smoothing state periodicity.
space
Neural Network Prediction by mathematical model High High Medium Yes Difficult
imitating the human brain.
Table 1: Representative forecast methods and characteristics
 Dynamic Ensemble Forecasting Technology:
Fujitsu laboratories have developed the original dynamic ensemble forecasting technology, which improves
the accuracy of demand forecasting and implements automatic tuning. This technology realizes both stable
forecasting accuracy for fluctuating demand and reduction of operational losses by means of automatic
tuning to maintain accuracy during operation. This technology is realized by combining conventional time
series analysis technology with machine learning.

In the field of Machine learning, a method called “ensemble learning” that realizes highly accurate
forecasting by combining many learning models is widely used. In the field of time series analysis, a
technique called “Model Integration” for calculating forecasted values by weighted averaging of forecasting
results produced by multiple methods is known as a related technique of ensemble learning.

It is difficult to select optimum method according to the forecasting target, and the forecasting condition is
fulfilled by machine learning and statistical methods. “Model selection” is often used as a way to solve this
challenge. In the case of forecasting target and forecasting conditions being fixed, the optimum forecast
method is selected automatically by this technology, according to the characteristics of each product. On the
other hand, when forecasting target varies according to the life cycle of the product, the optimum method
also changes dynamically accordingly.

The forecast error of model integration is smaller than that of each of the individual forecast methods used
for integration and forecasting accuracy is higher.

 Attribute Decomposition Model Technology:

Fujitsu laboratories developed an attribute decomposition model that forecasts initial demand from
planning information of new products by combining text mining and machine learning.

Demand forecasting for new products cannot be solved by Dynamic Ensemble forecasting technology alone,
because there is no past sales and shipment data available that can be used as a learning data. Hence,
Attribute decomposition model solves this problem by utilizing past results data of similar products.

The Attribute decomposition model applies text mining technology to extract and classify the attribute
information such as target layer and product characteristics of each product. Next, using Machine learning
technology, the sales contribution degree of each attribute information is quantitatively modelled from past
sales data. In forecasting demand for new products, sales volume is estimated by combining the attribute
information of the target produced with its sales contribution degree.

In the attribute decomposition model, information of past products is decomposed into product attributes
and sales contribution degree for use. There may be products that are partially similar to a new product, in
many cases no similar products overall exist. In this case, highly accurate forecasting of demand is possible
by using partially similar information with the help of attribute decomposition model.

 Applications & Expected effects:

FUJITSU offers ODMA (Operational Data Management & Analytics) Demand Forecasting SaaS (Software as a
service), a demand forecasting solution utilizing the dynamic ensemble forecasting technology and attribute
decomposition model technology as a cloud service from April 2018.

ODMA Demand Forecasting SaaS is a cloud service that collects performance data of organisations and other
characteristics data that affect demand, and forecasts future demand for an enterprise.
This solution provides a mechanism for the flexible creation and operation of forecasting models.
This solution was also designed for continuous growth of forecasting models.

At mass merchandisers with a large number of unskilled labours, ordering accuracy varies greatly
among individuals and order placement takes a long time. This problem can be solved by
calculating the optimum order quantities from the demand forecast and inventory situation as
recommended values. It also optimizes inventories to reduce sales opportunity losses due to
inventory losses as well as reduces unsold products.
In case of new products and seasonal goods, accurate demand forecasting is difficult because of the
variety of products and needs, and production plans are prepared using forecasting models based
on similar products. The forecasted demands often vary significantly from the actual demand,
resulting in disposal losses or sales opportunity losses.

This can be remedied by forecasting using planning information of new products and past sales and
shipment data of similar products to increase the forecasting accuracy of the demand and reduce
losses through effective product planning.

REFERENCE
I. FUJITSU Laboratories (Isamu Watanabe, Tooru Yoshibayashi & Tateki Imaoka)
“AI-Based Demand Forecasting for Both Reliable Forecasting and Efficient Operation:
Dynamic Ensemble Forecasting”
 A Review on Machine Learning-Based Demand Forecasting in Supply Chains
 Dhruv Sharma (2K16/AE/029)

INTRODUCTION
 A major facet of businesses today is the notion of supply chain integration, whereby resources are combined
to provide value to the end consumer and where all the upstream firms realize the importance of
integration. Such integration often relies heavily on, or at very least includes, sharing information between
various business partners Although integration and sharing information can potentially reduce forecast
errors, they are neither ubiquitous nor complete activities and forecast errors still abound. Collaborative
forecasting and replenishment (CFAR) permits a firm and its supplier-firm to coordinate decisions.

 Researchers have identified several factors that could hinder such long-term stable collaborative efforts.
Prem kumar (2000) lists some required critical issues that permit successful supply chain collaboration,
including:
(i) Alignment of business interests
(ii) Long-term relationship management
(iii) Reluctance to share information
(iv) Complexity of large-scale supply chain management
(v) Competence of personnel supporting supply chain management
(vi) Performance measurement and incentive systems to support supply chain management.

 The introduction of inaccurate information into the system could also lead to demand distortion.
The over-emphasis on investing in extensive relationships among the partners could lead to a “lock-
in” situation in their study they have found that while the quality of information sharing in a supply
chain could promote the flexibility, the breadth of information shared has a detrimental effect on it.
The modularity and loose couplings between the partners have been identified as positive factors in
this regard. In light of these considerations, the problem of forecasting distorted demand is of
significant importance to businesses, especially those operating towards the upstream end of the
extended supply chain.

 The purpose of this work is to investigate the potential value of applying advanced machine learning
techniques, including artificial neural networks (ANN), recurrent neural networks (RNN), and
support vector machines (SVM) to demand forecasting in supply chains. The performance of these
machine-learning (ML) methods is compared against baseline traditional approaches, such as
exponential smoothing, moving average, linear regression, and the Theta model.

LITERATURE REVIEW

 BACKGROUND

I. Demand Distortion in supply chains:

 One of the major purposes of supply chain collaboration is improving the accuracy of forecasts. The source
of the demand distortion in the extended supply chain simulation is due to demand signal processing by the
members in the supply chain (Forrester, 1961). According to Lee et al. (1997), demand signal processing
means that each party in the supply chain does some processing on the demand signal, thus transforming it
before passing it along to the next member. As the end-customer’s demand signal moves up the supply
chain, it becomes increasingly distorted. This occurs even if the demand signal processing function is
identical in all parties of the extended supply chain.

II. Traditional Forecasting techniques:

 Traditional forecasting, using historical observations estimate future business metrics such as inventory
requirements, asset performance, budgets and revenue. This practice fails to deliver the precision and agility
required of leading enterprises and effective asset management programs meeting difficult challenges. The
 reason for this shortfall is simple: the past does not represent the future in dynamic operations that must
respond to changing maintenance practices, aging equipment, engineering innovations, reliability
improvements, and highly variable world-wide operations.

III. Machine learning techniques:

 As the number of forecasting techniques and parameters increase, it becomes more difficult to choose the
appropriate one in a particular context. One possible solution is to rely on a class of algorithms called
“universal approximators,” which can approximate any function to an arbitrary accuracy. Using such
universal approximators, any required function between past and future data can be learned, thus
effectively making other forecasting techniques a subset of the functions that the universal approximators
can learn. ML techniques, such as artificial neural networks and support vector machines are universal
approximators and can be used to learn any function.

IV. Research Methodology :

 In essence, we set out to investigate whether ML, in general, performs better than traditional forecasting
techniques. To this end, we conducted experiments to compare the accuracy of ML forecasting techniques
with traditional ones in the context of noisy supply chain. In our study the traditional forecasting techniques
were represented by moving average, trend, exponential smoothing, and multiple linear regressions.
Research methodology is the specific procedures or techniques used to identify, select, process, and analyze
information about a topic. In a research paper, the methodology section allows the reader to critically
evaluate a study's overall validity and reliability.

 EXPERIMENTS

I. Data set preparation :

 We used a representative set of traditional forecasting techniques as a control group, and a set of machine
learning techniques as a treatment group. To compare the two groups, every technique from each group
was used to forecast demand one month into the future for all of the 100 series for the three datasets
previously identified. This resulted in a series of 4,700 forecast points for the chocolate manufacturer, 6,500
for the toner cartridge manufacturer and 14,800 for the Statistics Canada dataset for every technique tested.
However, since all forecasting techniques require past data to make a forecast into the future, there was a
predetermined start-up period that slightly reduced the number of forecast observations.

II. Models :
 The ARMA model combines both autoregressive forecast and a moving average forecast. To minimize the
error, we optimized the lag used in the auto-regression, IGI Global. Copying or distributing in print or
electronic forms without written permission of IGI Global is prohibited. Portion and the lag used in the
moving average portion. This functionality is provided by the MATLAB GARCH Toolbox. The Theta model in
our implementation mimicked the one used in M3 forecasting competition. First, the linear trend was
calculated, and then the exponential smoothing performed on double of the difference between the raw
data and trend values to minimize the error on the training set. The two individual series, the linear trend
and the optimized exponential smoothing on the decomposed series were recombined by an average of the
two.

CONCLUSION

 The purpose of this work has been to investigate the applicability and benefits of machine learning
techniques in forecasting distorted demand signal with a high noise in the context of supply chains. Although
there are several forecasting algorithms available to practitioners, there are very few objective and
reproducible guidelines regarding which method should be employed. In this research, we have shown
empirically that the best traditional method for a manufacturer is the automatic exponential smoothing with
the first value of the series as the initial value. We have also found that all of the more advanced machine
learning techniques have relatively poor performance, possibly due to the limited number of past time
periods for any given product. None of the ML techniques can reliably outperform the best traditional
counterpart (exponential smoothing) when learning and forecasting single time series.
 However, one important finding concerns the usefulness of combining the data about multiple products in
what we called a super wide model in conjunction with a relatively new technique, the support vector
machine (SVM). The domain-specific empirical results show that this approach is superior to the exponential
smoothing. The error reduction found range from 3.11 percent to 10 percent, which can result in large
financial savings for a company depending on the cost related to inventory errors.

 Since the work used actual data from a large number of products from two North American manufacturers,
with the additional verification against Statistics Canada manufacturing survey. Number of products added
to the combined time series model (super wide approach) increases, the performance could also increase.

 One important point to note is that SVM are computationally intensive and the cross validation based
complexity parameter optimization procedure results in running a large amount of support vector machines
depending on the precision of the complexity search. The longest running models in this research took over
3 days of processing on a modern computer. There are many optimization techniques that could be
performed to reduce the processing time such as parallelization, which is trivial for a cross-validation
procedure and reduction of the complexity term search precision. Further optimizations to the SVM
algorithms and the increase in processing power would reduce the processing time significantly. Once the
models have been completed, they can be used for forecasting with relatively little processing time.
 A Review on Machine Learning Methodology Applied to
Demand and Sales Forecasting
 Abbimnyou Warikoo (2K16/AE/001)

INTRODUCTION
The primary target of any Supply Chain (SC) is to maximize the total generated value, which is the difference
between what the client pays for the final product and the costs incurred by the chain in order to fulfil the client’s
order. Acing SC gives an immense upper hand, as it enables an organization to boost this created worth. This is the
motivation behind why most by far of organizations attempt to be at the bleeding edge of forefront patterns.
Because of progress in the course of recent decades in registering power (huge information stockpiling, even more
dominant processors, quicker web association, and so forth.) issues that appeared to be very perplexing or costly to
tackle are presently inside our scope. New patterns, for example, Big Data, Cyber security, Internet of Things (IoT)
and Crypto currencies have risen, prepared to challenge the previously mentioned innovative leaps forward. For
example, IoT, which intends to utilize inserted frameworks including sensors and actuators combined with the web
to empower control and quick access to data progressively, speaks to one of these difficulties, as there will be
around 25-50 billion associated gadgets by 2020.
One of the most recent and popular techniques that aims to tackle those new business challenges is Big Data
Analytics (BDA). An accurate definition of BDA is, “the application of advanced analytic techniques including data
mining, statistical analysis, predictive analytics, etc. on big data sets as new business intelligence practice”. In short,
it is the coordination of Big Data and Machine Learning (ML) techniques to provide trustworthy insights for decision
making. The general architecture and elements in a BDA model. Two main ideas can be extracted from this model:
firstly, ML and big data benefit from each other, as they can be coupled to create more complete models. Secondly,
the main purpose of BDA is to transform information into useful knowledge. In this context, the objective of this
paper is to provide an insight about the most recent ML techniques applied to D&SF by means of a literature review.

LITERATURE REVIEW
1. MACHINE LEARNING

People learn through understanding: we utilize an experimentation procedure to find which activities ought
to be activated given certain conditions. This enables us to make reflections and assemble information. AI is
some way or another comparative; it tends to be viewed as calculations having for goal to improve an
exhibition measure via naturally inferring its own principles and making its very own models dependent on
given data. In general terms, we can identify three types of learning methods: supervised learning,
unsupervised learning and reinforcement learning.

1.1 Supervises Learning:

This kind of learning is characterized by using data structures that have a collection of features pointing
to a result (output). As the desired output is already known, supervised models learn progressively to
imitate the required output. In that purpose, the learning system creates its own logic, that may produce
outputs to queries on new features. Supervised Learning is often used for classification and regression.

In classification, the objective is to predict the categorical class label of new instances by using rules
learnt from past observations. Classification may be binary or may concern multiple classes. In binary
classification, the model learns a series of rules that will distinguish between two possibilities
(i.e. predicted stock trends by classifying whether the price will increase or decrease using Decision
Trees and Support Vector Machine); while multiclass classification tries to map a new instance onto one
 of several possibilities. For instance, differentiating between wine samples coming from three different
growers by knowing their intrinsic features such as malic acid, alkalinity of ash, magnesium level, etc.

1.2 Unsupervised Learning:

In contrast to the learning type, an unsupervised learning system is not fed with an expected output or
explicit feedback to create its rules; instead, the system is supposed to uncover patterns by exploring a
data structure and extracting meaningful information. This type of learning is for instance used for data
clustering, which is a way of organizing a data set into subgroups in order to recognize patterns or
hidden structures. It is also used in dimensionality reduction, which consists in choosing the most
meaningful variables for the model. The latter application is convenient when models are complex and
computationally expensive, as it reduces the number of features considered.

1.3 Reinforcement Learning:

Reinforcement learning is related to Supervised Learning, but “instead of training examples that indicate
the correct output for a given input, the training data in reinforcement learning are assume to provide
only an indication as to whether an action is correct or not (...)”. In short, the objective is to maximize
the expected reward over time by learning a series of actions that will avoid punishments or penalties.
One of the most impressive and most recent applications of reinforcement learning is AlphaGo, the
Artificial Intelligence

Program that beat the European champion Fan Hui 5-0 at the Chinese board game go in October 2015. A
new version of Alpha Go, Alpha Go Zero, was released in 2017. Unlike its earlier versions, Alpha Go Zero
only learned by Reinforcement Learning and the results were astonishing: Alpha Go Zero Beat Alpha Go
100-0.

2. MACHINE LEARNING IN DEAMAND AND SALES FORCASTING

Before tending to the theme, explain that Demand Forecasting is unique in relation to Sales Forecasting. The
last uses information that have been legitimately gathered from the POS (Point of Sale), so it is dependent
upon the impact of advancements or stock deficiencies. Then again, Demand Forecasting utilizes information
where the impact of advancements or deficiencies has been adjusted, with the goal of mirroring the genuine
market request. Regardless of this inconspicuous yet significant distinction, this paper will cover the two
ideas under a solitary term: Demand and Sales Forecasting (D&SF).

D&SF is one of the most important elements of any SC. It aims to coordinate all the SC partners by reducing
the information lag between the final client and the n-tier supplier. This provides several advantages, such as
mitigation of the bullwhip effect, better resources and aggregate capacity planning, and reduction of stock
shortages or oversupply costs. The most widely known statistical techniques meant to forecast are also
called traditional forecasting methods.

Traditional forecasting methods are based on time series. This means they are applied under the hypothesis
that the past demand can statistically estimate the future demand. Normally, these methods are easy to
apply and present good performance in markets whose demand is mostly stable. Unfortunately, this is not
often the case: the demand also depends on exogenous factors that are not effectively represented by past
values. For instance, on-demand ride services like UBER, Lyft or Didi Chuxing cannot estimate their demand
by only relying on time series, they must consider other elements like weather conditions (such as humidity,
temperature, etc.), time of the day or day of the week.

2.1 Techniques in literature:

These days, ML profits by a decent notoriety, likely because the greater part of the previously
mentioned requirements have been survived. On account of this, new applications have been
distributed prompting new patterns and methods. So as to recognize these new patterns and meet our
examination objective, a writing study was performed by reviewing three databases (Science Direct,
Scopus and IEEE) with the accompanying arrangements of watchwords in titles, modified works, and
 catchphrases: 'AI' and 'Inventory network', 'AI' and 'Store network Management', 'Neural Networks' and
'Determining' and 'Production network', 'Huge Data' and 'AI' and 'Anticipating', 'AI' and 'Deals
Forecasting', 'AI' and 'Request Forecasting'. Thereafter, just papers tending to Demand or Sales gauging
themes and distributed somewhere in the range of 2009 and 2017 were preselected. At long last, from
the preselected papers, a short rundown was worked by picking the most relevant articles for
examination: late papers (distributed after 2015) were possibly dissected on the off chance that they
were referred to at any rate once, while papers going before 2015 were considered on the off chance
that they were much of the time referred to.

The short rundown was made from 10 distributions. For each investigation, the application, the sort of
highlights found in the dataset, the strategy used for information pre-processing, and the utilized ML
strategies for D&SF are introduced.

2.2 Comparison between ML and classical forecasting methods:

Traditional models offer huge advantages in terms of simplicity and accuracy, as they can perform D&SF
in a matter of seconds for several SKUs (Stock Keeping Units). Nevertheless, they need to be thoroughly
designed by an expert who can fit the precise needs of the enterprise to a specific model. Moreover,
they fail to include exogenous variables. ML models partially solve this problem because they can
include other types of data like endogenous and exogenous variables, allowing for a better
representation of reality. What is more, ML techniques that are properly implemented outperform most
of the traditional forecasting methods.
The fact that ML outperforms traditional models at D&SF does not necessarily mean that companies
should change of their forecasting tools. This leads us to consider an important question: when should
enterprises invest in ML for D&SF?

D&SF must be flexible and reactive, especially when dealing with short-term forecasting, which is
necessary in sectors such as the clothing industry. The calculation of the forecast must be agile, because
most companies have hundreds, thousands or even more SKUs, and all of them may need an immediate
estimation. If not applied appropriately, ML models like ANN could take hours or even days to train.
Moreover, after the training phase, these methods are still time-consuming, making them less suitable
for D&SF. For instance, [9] applied both SVM and the classical Linear Regression using all the features.
The former took around 6000 seconds to predict sales, whereas the latter only took around 980
seconds. In this case, SVM would not be convenient for applications dealing with several case scenarios,
as the relation between accuracy and computing time is not attractive.

CONCLUSION
Managing a high volume of information is one of the more typical difficulties forced by present day business
patterns. Since information has gotten one of the most significant assets, SC directors are anxious to extricate
pertinent data that will prompt an upper hand. This paper utilized a writing concentrate to examine the utilization of
ML in D&SF as a technique to accomplishing this preferred position. Ten late explore papers applying ML to D&SF
were chosen and broke down to distinguish new drifts. One of the discoveries was that ML widens the range of
D&SF, as it can deal with complex factors. More absolutely, Fuzzy-ANN approaches indicated astounding execution
when managing uncertain information like climate factors, while DT and RF offered significant translation limit.
Besides, Data Pre-processing methods demonstrated to considerably decrease the multifaceted nature of the
models, empowering both great precision and sensible processing time.
At this stage, it very well may be seen that no examination utilizing information mining procedures on D&SF was
discovered, which is somewhat amazing. Systems, for example, rule mining may permit to examine a lot of
information and to distinguish relationships between things that can give unequivocal information, ready to improve
the abilities of the chiefs. This might be of explicit enthusiasm for better formalizing the job of the exogenous
factors, and this office can be utilized in blend with conventional prescient techniques.
As adjusting an organization to new innovations frequently accompanies questions raised by vulnerability, a
correlation among ML and customary gauging strategies was made with the goal of giving knowledge to chiefs who
are eager to execute ML in their procedures. The after-effects of this examination show that ML is a more
reasonable strategy than conventional anticipating strategies as far as exactness, especially when models contain
exogenous and endogenous factors.

REFERENCE
I. Juan Pablo Usuga Cadavid, Samie Lamouri & Bernard Grabot “Trends in Machine Learning Applied to
Demand & Sales Forecasting: A Review”
A Review on Application of machine learning
techniques for supply chain demand forecasting
 Arnab Biswas (2K16/AE/017)

INTRODUCTION
Demand Forecasting is the process in which historical sales data is used to develop an estimate of an
expected forecast of customer demand. The objectives of this research are to study the feasibility and perform a
comparative analysis of forecasting the distorted demand signals in the extended supply chain using non-linear
machine learning techniques.

Non- linear machine learning is used because when traditional methods are used the firm’s demand appears to
fluctuate in a random fashion even if the final customer’s demand has a predictable pattern, this know as Bullwhip
effect. Forecasting under these conditions becomes very difficult. That is why machine learning in supply chain
demand forecasting is useful tool.

The supply chain demand forecasting methods which are used in the research are -

 TRADITIONAL FORECASTING METHODS:

1. Naive Forecast- It simply uses the latest value of the variable of interest as a best guess for the future
value.

2. Average- A number representing the central or typical value in the set.

3. Moving average- Uses the average of a defined number of previous periods as the future forecasted
demand.

4. Trend- based on a simple regression model that takes time as an independent variable and tries to forecast
demand as a function of time.

5. Multiple Linear Regressions - tries to predict the change in demand using a number of past changes in
demand observations as independent variables.

 ADVANCED FORECASTING METHODS:

1. Neural Networks- In these networks, the individual elements (‘‘neurons’’) are organized into layers in
such a way that output signals from the neurons of a given layer are passed to all of the neurons of the
next layer. Thus, the flow of neural activations goes in one direction only, layer-by-layer.

2. Recurrent Neural Networks- Recurrent neural networks allow output signals of some of their neurons
to flow back and serve as inputs for the neurons of the same layer or those of the previous layers.

3. Support vector machines- are a newer type of universal function approximators that are based on the
structural risk minimization principle from statistical learning theory

LITERATURE REVIEW
  In order to examine the effectiveness of various advanced machine learning techniques in forecasting supply
chain demand, two data sets were prepared. The first one represents results of the simulations of an
extended supply chain, and the second one contains actual Foundries data provided by Statistics Canada.

 The simulation includes four parties that make up the supply chain that delivers a product to a final customer.
Demand signal processing is modelled by a simple linear regression that calculates the trend over the past 10
days, and, which is then used to forecast the demand in 2 days.

 The mechanism for demand signal processing is represented diagrammatically in Fig. As can be noted, the
demand signal, after having been processed by the demand forecasting function, feeds into the purchase-
order calculation function. The Delivery, Inventory and Backlog signals are all part of the simulated partner
model in order to enable simulation monitoring and data extraction.

Mechanism for Demand Signal Processing

 The result of the experiment shows that support vector machines have the best results. Their accuracy is
even higher in Foundries demand data series than in the stimulated data series taken.

 Moving average, naive and trend forecasting have been among the worst performers.

 However, statistical analysis of the results shows that there was no statistically significant difference in terms
of the accuracy of forecasts among RNN, SVM, NN, and MLR.

CONCLUSION
 Ability to increase forecasting accuracy results in lower costs and higher customer satisfaction because of
more on-time deliveries.

 The advanced techniques did not provide a large improvement over more ‘‘traditional techniques’’ (as
represented by the MLR model) for the simulation data set. However, for the real foundries data, the more
 advanced data mining techniques (RNN and SVM) provide larger improvements. Recurrent Neural Networks
(RNN) and Support Vector Machines (SVM) provide the best results on the foundries test set. We can also
see that the trend estimation and naive forecast are the worst types of demand signal processing since they
have the highest level of error.

 Machine learning techniques and MLR for forecasting distorted demand signals in the extended supply chain
provide more accurate forecasts than simpler forecasting techniques (including naive, trend, and moving
average).

REFERENCE
I. Real Carbonneau, Kevin Laframboise & Rustam Vahidov “Application of Machine
Learning techniques for supply chain demand forecasting”.
 CONCLUSION
In supply chain distribution industry, demand forecasting is one of the main problems of supply chains to optimize
stocks, reduce costs, and increase sales, profit, and customer loyalty. To overcome this issue, there are several
methods such as time series analysis and machine learning approaches to analyze and learn complex interactions
and patterns from historical data.

In this report, there is a novel attempt to integrate different forecasting models that include time series algorithms,
linear regression models, deep learning method and machine learning approaches such as dynamic ensemble
learning technology and attribute decomposition model technology for demand forecasting process. The proposed
forecasting technologies are tested and carried out real life data from different organisations. It is observed that the
inclusion of various learning algorithms advanced the performance of demand forecasting system.

This paper introduced the dynamic ensemble forecasting technology and attribute decomposition model technology
developed by Fujitsu Laboratories to provide solutions to the challenge of demand forecasting in the distribution
industry. It also described solutions based on these technologies and actual application cases.

This review also investigates the acceptability and benefits of machine learning techniques in forecasting distorted
demand signal with a high noise in the context of supply chains. The error reduction found range from 3.11% to 10%,
which can result in large financial savings for a company depending on the cost related to inventory errors. Although
showing better results overall, the advanced techniques did not provide a large improvement over more “traditional
techniques”.

In this paper, a comparison between Machine Learning and Traditional forecasting methods was made with the
objective of providing insight to managers who are willing to implement ML in their processes. The results of this
study show that ML is a more suitable technique than traditional forecasting methods in terms of accuracy,
particularly when models contain exogenous and endogenous variables. Furthermore, it enables the identification of
hidden patterns in demand that can be used as a baseline to identify new market trends.

Overall, we can conclude that the use of machine learning techniques for forecasting demand in the supply chain
provide more accurate forecasts than simpler or traditional forecasting techniques. However, we did not find that
machine learning techniques show significantly better performance than linear regression.

Machine Learning Forecasting is the future of demand administration

 FUTURE SCOPE OF MACHINE LEARNING IN
DEMAND FORECASTING
Supply chain experts have, for long, relied on the traditional forecasting models and tools to derive demand forecast
for their goods and services. While it has helped them reduce the demand-supply mismatch to some extent, there’s
lot of catching up to do to improve the forecasting models.

Thanks to machine learning, the demand planners now have the right set of tools and algorithms to deliver available
to promise which in turn will bring about remarkable improvements to customer satisfaction, revenue gains and
inventory cost optimization.

Various organisations are working o improving demand forecasting for a number of industries and business types,
including the distribution industry, based on the demand forecasting technologies and solutions. In addition to
further refinement of the forecasting technologies, the advancement of demand forecasting requires an approach
that expands the data used as predictors to allow more accurate gasp and forecasting of demand.

One important possibility for future research could be investigating the benefits of Machine Learning techniques
when using other sources of data and in the context of collaborative forecasting. This additional data may include
economic indicators, market indicators, collaborative information sources, product group averages and other
relevant information. This additional information would likely increase the model’s accuracy.

For further research, Data Preprocessing techniques should be studied, as they offer significant advantages in terms
of reduction of computational costs when applying Machine Learning models. Finally, addressing new trends to
SMEs is vital, as they are numerous but often lack the financial means or experience to implement cutting-edge
technologies.

Future research could be directed to investigating the impacts of information sharing on forecasting accuracy, e.g.,
using the Internet, and other e-business technologies, as a means of enabling firms to coordinate decisions with their
various partners.

With the levels of incremental improvements that Machine Learning can drive, companies should augment their as-
is demand forecasting techniques with machine learning powered models. It’s time to look beyond traditional ways
of demand forecasting and embrace digital to minimize the demand-supply gap as much as possible. The off the
shelf products will never be able to match the capabilities of machine learning models and there’s no other way out
but to embrace this change soonest.