A comparative study of

time series forecasting

methods using real data
Submitted for the partial fulfilment of the degree of
Master of Business Administration

SANDEEP SHAH (ROLL NO - 1753817)

MBA – FINAL YEAR
(DEI, SURAT)

DEI | Dayalbagh, Agra
 Page |1

Table of Contents

Abstract............................................................................................................................................3
Chapter 1: Introduction....................................................................................................................4
Executive Summary....................................................................................................................4
Research Design........................................................................................................................10
Chapter 2: Review of Literature....................................................................................................12
Study of Literatures, Articles and Research Papers................................................................12
Need of study.............................................................................................................................13
Chapter 3: Data Analysis & Interpretation....................................................................................14
Brief introduction to time series forecasting models...............................................................14
Data Analysis and Modeling.....................................................................................................15
Product: GALCF........................................................................................................................18
Chapter 4: Findings, Recommendations and Conclusions............................................................22
Findings.....................................................................................................................................22
Conclusions...............................................................................................................................22
References and Bibliography.........................................................................................................23
References.................................................................................................................................23
Bibliography..............................................................................................................................23
 Page |2

Acknowledgment
I would like to acknowledge the kind assistance of all the persons who have contributed directly
or indirectly in completing this thesis.
RAHUL KUMAR, CIO, AMNS India, who first sparked my interest in Artificial Intelligence,
for his critical, yet always supportive, supervision, and for giving me the opportunity to look into
a very interesting topic.
All my fellows and colleagues have provided a good support in data collection and converting
the ideas into a report.
I am thankful to the Faculty Coordinator MR. SAURAV KUMAR for his support to undertake
the study and to complete this in time. I would also like to thank the researchers and the scholars
to whom I concerned with by using their project reports and findings is. I am very grateful to the
executives and sales representatives who have helped in understanding the data and supported to
perform such analysis.
 Page |3

Abstract
This report is addressed to all readers interested in harnessing the power of data analysis,
machine learning and other statistical techniques for a better planning like sales forecasting that
helps in better plan their production, stocks and other important consideration. The aim of this
dissertation is to study the various time series techniques for forecasting and hen implementing
the most suitable technique to forecast the sales of products of AMNS India for next financial
year.
A forecast is a prediction or estimation of future situation. It is an objective assessment of future
course of action. Since future is uncertain, no forecast can be cent per cent correct. Forecasts can
be both physical as well as financial in nature. The more realistic the forecasts, the more
effective decisions can be taken for tomorrow.
The method and technique will be based on research reports on forecasting techniques, use cases
and articles from technical or data science papers, blogs etc. to understand the forecasting
technique and their pros and cons.
In this paper a comparative study of two commonly used linear demand models for
statistical prediction is presented. The models are Autoregressive Integrated Moving Average
(ARIMA) model and Holt-Winters model. These models have exceptional adaptive capacity
to deal with the linearity in problem solving . The aim of the study is to find the model that
provides more accurate prediction. The comparison is made by considering the smallest values
of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), for prediction based on
daily, weekly and monthly time series.
For implementation part the primary data will be collected from the data ware house of the
company for last 3 financial year.
Thus, this report is subject to a comparative analysis of various techniques available for time
series data and will be used for real data to see how effectively it can give estimation of sales for
next upcoming financial year.
 Page |4

Chapter 1: Introduction

Executive Summary
In this age of cut throat competition, everyone is struggling to gain a slice of bread in highly
VUCA world where consumers are highly unpredictable and market conditions are very
uncertain.
Time series analysis comprises methods for analyzing time series data in order to extract
meaningful statistics and other characteristics of the data. Time series forecasting is the use of a
model to predict future values based on previously observed values.
Time series are widely used for non-stationary data, like economic, weather, stock price, and
retail sales. There is a management philosophy given by Winston Churchill is credited with
saying: “He who fails to plan  is planning to fail”.
In this report discussion will be on why conventional techniques are not so effective and why do
we need advanced data driven techniques to solve the complex trends and find the close
estimation for demand of the product.
Usually, company makes sales forecast based on internal as well as external data but due to
limitation of resources and time, we will consider only the company demand data and based on
advanced statistical analysis will predict the demand for next financial year.

Background of the Problem

Marketing professionals face increasing complexity due to the uncertain market conditions and
unpredictable customer behavior and it hampers to have a better planning for the future demand
followed by resource and raw material planning. Especially in the case of heavy industries like
steel and mining it is very difficult to manage when the estimated demand in not there
beforehand and leads to huge loss due to unplanned stock and inventory.
All requirements of the business sector need the technique of accurate and practical reading into
the future. Forecasts are, therefore, very essential requirement for the survival of business. Man-
agement requires forecasting information when making a wide range of decisions.

The sales forecast is particularly important as it is the foundation upon which all company plans
are built in terms of markets and revenue. Management would be a simple matter if business was
not in a continual state of change, the pace of which has quickened in recent years.

It is becoming increasingly important and necessary for business to predict their future prospects
in terms of sales, cost and profits. The value of future sales is crucial as it affects costs profits, so
the prediction of future sales is the logical starting point of all business planning.
In this report, the demand forecasting will be done for AMNS India for last 3 years’ demand
data.
 Page |5

Basic concepts and terminologies related to demand forecasting

Meaning of Demand Forecasting:

Forecasts are becoming the lifetime of business in a world, where the tidal waves of change are
sweeping the most established of structures, inherited by human society. Commerce just happens
to the one of the first casualties. Survival in this age of economic predators, requires the tact,
talent and technique of predicting the future. Forecast is becoming the sign of survival and the
language of business.and Neural Network. The purpose of this paper is not to dive deep into these
algorithms instead to focus on how these can be used for growth in sales for business.

In the words of Cundiff and Still, “Demand forecasting is an estimate of sales during a specified
future period which is tied to a proposed marketing plan and which assumes a particular set of
uncontrollable and competitive forces”. Therefore, demand forecasting is a projection of firm’s
expected level of sales based on a chosen marketing plan and environment.

Procedure to Prepare Sales Forecast:

Companies commonly use a three-stage procedure to prepare a sales forecast. They make an
environmental forecast, followed by an industry forecast, and followed by a company’s sales
forecast, the environmental forecast calls for projecting inflation, unemployment, interest rate,
consumer spending, and saving, business investment, government expenditure, net exports and
other environmental magnitudes and events of importance to the company.

The industry forecast is based on surveys of consumers’ intention and analysis of statistical
trends is made available by trade associations or chamber of commerce. It can give indication to
a firm regarding the direction in which the whole industry will be moving. The company derives
its sales forecast by assuming that it will win a certain market share. All forecasts are built on
one of the three information bases: What people say? What people do? What people have done?

Types of Demand Forecasting:

Forecasts can be broadly classified into:

(i) Passive Forecast, and

(ii) Active Forecast.

Under passive forecast prediction about future is based on the assumption that the firm does not
change the course of its action. Under active forecast, prediction is done under the condition of
likely future changes in the actions by the firms.
 Page |6

From the view point of ‘time span’, forecasting may be classified into two, viz.
(i) Short term demand forecasting and
(ii) Long term demand forecasting.
In a short run forecast, seasonal patterns are of much importance. It may cover a period of three
months, six months or one year. It is one which provides information for tactical decisions.

Which period is chosen depends upon the nature of business. Such a forecast helps in preparing
suitable sales policy. Long term forecasts are helpful in suitable capital planning. It is one which
provides information for major strategic decisions. It helps in saving the wastages in material,
man hours, machine time and capacity. Planning of a new unit must start with an analysis of the
long term demand potential of the products of the firm.

There are basically two types of forecast, viz.

(i) External or national group of forecast, and
(ii) Internal or company group forecast. External forecast deals with trends in general
business. It is usually prepared by a company’s research wing or by outside
consultants.

Internal forecast includes all those that are related to the operation of a particular enterprise such
as sales group, production group, and financial group. The structure of internal forecast includes
forecast of annual sales, forecast of products cost, forecast of operating profit, forecast of taxable
income, forecast of cash resources, forecast of the number of employees, etc.

At different levels forecasting may be classified into

(i) Macro-level forecasting,
(ii) Industry- level forecasting,
(iii) Firm- level forecasting and
(iv) Product-line forecasting. Macro-level forecasting is concerned with business
conditions over the whole economy. It is measured by an appropriate index of
industrial production, national income or expenditure.

Industry-level forecasting is prepared by different trade associations. This is based on survey of

consumers’ intention and analysis of statistical trends. Firm-level forecasting is related to an
individual firm. It is most important from managerial view point. Product-line forecasting helps
the firm to decide which of the product or products should have priority in the allocation of
firm’s limited resources.

Forecast may be classified into

(i) General and
(ii) Specific.
The general forecast may generally be useful to the firm. Many firms require separate forecasts
for specific products and specific areas, for this general forecast is broken down into specific
forecasts.

There are different forecasts for different types of products like

(i) Forecasting demand for nondurable consumer goods,
 Page |7

(ii) Forecasting demand for durable consumer goods,

(iii) Forecasting demand for capital goods, and
(iv) Forecasting demand for new-products.

In this, report focus is on capital goods and that is also for a B2B customers.

Forecasting Demand for Capital Goods:

Capital goods are used for further production. The demand for capital good is a derived one. It
will depend upon the profitability of industries. The demand for capital goods is a case of
derived demand. In the case of particular capital goods, demand will depend on the specific
markets they serve and the end uses for which they are bought.

The demand for textile machinery will, for instance, be determined by the expansion of textile
industry in terms of new units and replacement of existing machinery. Estimation of new
demand as well as replacement demand is thus necessary.

Three types of data are required in estimating the demand for capital goods:

(a) The growth prospects of the user industries must be known,

(b) The norm of consumption of the capital goods per unit of each end-use product must be
known,
(c) The velocity of their use.

Techniques of Demand Forecasting:

Demand forecasting is a difficult exercise. Making estimates for future under the changing con-
ditions is a Herculean task. Consumers’ behaviour is the most unpredictable one because it is
motivated and influenced by a multiplicity of forces. There is no easy method or a simple
formula which enables the manager to predict the future. Economists and statisticians have
developed several methods of demand forecasting.
Demand forecasting is a difficult exercise. Making estimates for future under the changing con-
ditions is a Herculean task. Consumers’ behaviour is the most unpredictable one because it is
motivated and influenced by a multiplicity of forces. There is no easy method or a simple
formula which enables the manager to predict the future. Economists and statisticians have
developed several methods of demand forecasting.
Each of these methods has its relative advantages and disadvantages. Selection of the right
method is essential to make demand forecasting accurate. In demand forecasting, a judicious
combination of statistical skill and rational judgment is needed.
 Page |8

Mathematical and statistical techniques are essential in classifying relationships and providing
techniques of analysis, but they are in no way an alternative for sound judgment. Sound
judgment is a prime requisite for good forecast.
The judgment should be based upon facts and the personal bias of the forecaster should not
prevail upon the facts. Therefore, a mid-way should be followed between mathematical
techniques and sound judgment or pure guess work. The more commonly used methods of
demand forecasting are discussed below.
The various methods of demand forecasting can be summarised in the form of a chart as shown
below:

It is difficult to assess each of the methodology in detail. So, for this project only Holt-Winter
Model will be focused and explained in detail.

Criteria of a Good Forecasting Method:

There are thus, a good many ways to make a guess about future sales. They show contrast in
cost, flexibility and the adequate skills and sophistication. Therefore, there is a problem of
choosing the best method for a particular demand situation. There are certain economic criteria
of broader applicability.

(i) Accuracy:

The forecast obtained must be accurate. How is an accurate forecast possible? To obtain an
accurate forecast, it is essential to check the accuracy of past forecasts against present
 Page |9

performance and of present forecasts against future performance. Accuracy cannot be tested by
precise measurement but buy judgment.

(ii) Plausibility:

The executive should have good understanding of the technique chosen and they should have
confidence in the techniques used. Understanding is also needed for a proper interpretation of
results. Plausibility requirements can often improve the accuracy of results.

(iii) Durability:

Unfortunately, a demand function fitted to past experience may back cost very greatly and still
fall apart in a short time as a forecaster. The durability of the forecasting power of a demand
function depends partly on the reasonableness and simplicity of functions fitted, but primarily on
the stability of the understanding relationships measured in the past. Of course, the importance of
durability determines the allowable cost of the forecast.

(iv) Flexibility:

Flexibility can be viewed as an alternative to generality. A long lasting function could be set up
in terms of basic natural forces and human motives. Even though fundamental, it would
nevertheless be hard to measure and thus not very useful. A set of variables whose co-efficient
could be adjusted from time to time to meet changing conditions in more practical way to
maintain intact the routine procedure of forecasting.

(v) Availability:

Immediate availability of data is a vital requirement and the search for reasonable
approximations to relevance in late data is a constant strain on the forecasters’ patience. The
techniques employed should be able to produce meaningful results quickly. Delay in result will
adversely affect the managerial decisions.

(vi) Economy:

Cost is a primary consideration which should be weighed against the importance of the forecasts
to the business operations. A question may arise: How much money and managerial effort should
be allocated to obtain a high level of forecasting accuracy? The criterion here is the economic
consideration.

(vii) Simplicity:

Statistical and econometric models are certainly useful but they are intolerably complex. To
those executives who have a fear of mathematics, these methods would appear to be Latin or
Greek. The procedure should, therefore, be simple and easy so that the management may
appreciate and understand why it has been adopted by the forecaster.
 P a g e | 10

(viii) Consistency:

The forecaster has to deal with various components which are independent. If he does not make
an adjustment in one component to bring it in line with a forecast of another, he would achieve a
whole which would appear consistent.
 P a g e | 11

Objectives of the study

i. Comparative study of various time series techniques available for demand/sales
forecasting: While forecasting is very important for a business then the selection of right
forecasting method is also crucial in order to move in a right direction. So, comparative
study is to be done to find the best suitable method for the selected company.

ii.Using the best suitable technique, prediction of the demand for the selected product
of the company. This is the main objective where the practical implementation on real
data to be used to predict the demand for next financial year for the company. The method
is to be selected based on the analysis of review of literature.

Research Design
Problem statement
Forecasting reduces the risk associated with business fluctuations which generally produce harm-
ful effects in business, create unemployment, induce speculation, discourage capital formation
and reduce the profit margin.

Forecasting is indispensable and it plays a very important part in the determination of various
policies. In modern times forecasting has been put on scientific footing so that the risks
associated with it have been considerably minimized and the chances of precision increased.

In most of the advanced countries there are specialized agencies. In India businessmen are not at
all interested in making scientific forecasts. They depend more on chance, luck and astrology.
They are highly superstitious and hence their forecasts are not correct.

Sufficient data are not available to make reliable forecasts. However, statistics alone do not
forecast future conditions. Judgment, experience and knowledge of the particular trade are also
necessary to make proper analysis and interpretation and to arrive at sound conclusions.

Now how to use internal data without considering the external data to make a closer
forecast is the challenging task that is being tried to resolve in this project.

Data Collection
Data will be collected from the data warehouse of the company for selected products for last 3
years starting from FY 2017 to FY 2019.
 P a g e | 12

Scope of Study
The study is both qualitative and quantitative. The data used for quantitative data analysis is for
selected company and for the selected period only. Also the forecasting will be made only on
quantity ordered without any other parameters. The technique used will be Time Series
Forecasting.

Tools to be used
Data Collection: Documents, Observation, Literature Reviews, Data warehouse
Data Analysis: Visual Tools, descriptive statistics, exploratory data analysis, time series
forecasting models i.e. Holt-Winter Exponential Method

Limitations of the study

The study has number of limitations.
i. Qualitative study of the methodologies for demand forecasting due to time constraint,
all methods cannot be tested
ii. Methods selected for forecasting is not completely unbiased due to the availability of
data, time and cost factors.
 P a g e | 13

Chapter 2: Review of Literature

Study of Literatures, Articles and Research Papers
S. No. Book/Journal/ Author Takeaway
Publisher/ (Year)
Title
1 Tehnički [ CITATI This article aims to compare the performances between
vjesnik 24, ON Güz16 Trend Analysis, Decomposition and Holt-Winters (HW)
Suppl. 2(2017), \l 16393 ] models for the prediction of a time series formed by a
503-509 group of jam and sherbet product demands.

Three forecasting methods were utilized and compared

Demand
with each other to find out the most successful method
Forecasting: A
within the given data set.
Comparison
Between The
Decomposition method has high performance in ‘cyclic’
Holt-Winters,
models, and HW method is successful in ‘seasonal’
Trend Analysis
models where cyclic data are also existent.
And
Decomposition
Models After the prediction process, the performances of all the
methods were compared and it was observed that the
Decomposition and the HW methods were both successful
in a trendy-seasonal-cyclic data like our group.
2 Researchgate.N [ CITATI In this paper, both Seasonal ARIMA and Holt-Winters
et ON models are developed to predict the monthly car sales in
Ale18 \l South Africa using data for the period of January 1994 to
Risk 16393 ] December 2013.
governance &
The purpose of this study is to choose an optimal model
control: suited for the sector. The three error metrics; mean
financial absolute error, mean absolute percentage error and root
markets & mean square error were used in making such a choice.
institutions /
Volume 6, Upon realizing that the three forecast errors could not
Issue 1, Winter provide concrete basis to make conclusion, the power test
2016 was calculated for each model proving Holt-Winters to
having about 0.3% more predictive power. Empirical
results also indicate that Holt-Winters model produced
more precise short-term seasonal forecasts.
3 Machinelearnin [CITATIO Classical methods like ETS and ARIMA out-perform
gmastery- N Jas18 \l machine learning and deep learning methods for one-
Comparing 16393 ] step forecasting on univariate datasets.
Classical and Classical methods like Theta and ARIMA out-
Machine perform machine learning and deep learning methods
Learning for multi-step forecasting on univariate datasets.
Algorithms for Machine learning and deep learning methods do not
Time Series yet deliver on their promise for univariate time series
Forecasting forecasting, and there is much work to do.
 P a g e | 14

4 http://ceur- [CITATIO The focus of this paper was on applying time series
ws.org/Vol- N Mic18 \l modeling techniques to a collection of USPTO patents
1353 16393 ] from 1996 to 2013. The techniques used are Holt-Winters
A Comparison Exponential Smoothing and ARIMA.
of Time Series
An interesting observation from the experimentation is
Model that the Database and Control System patent groups
Forecasting favored an ARIMA model, while Artificial Intelligence
Methods on patents fit better with a Holt Winters model.
Patent
Groups
5 researchgate.ne [ CITATI In this work, comparative study of two models for
t ON Mar18 statistical prediction of energy consumption is made.
\l 16393 ] These models are Autoregressive Integrated Moving
Comparative Average (ARIMA) model and Holt-Winters model.
study of
It is shown that ARIMA model gives more accurate
ARIMA and results when the prediction is based on daily time
Holt-Winters series data, while Holt-Winters model has better
statistical performance for weekly and monthly time series data.
models for
prediction of
energy
consumption

Need of study
This is clear that the forecasting is important for any organization but selection of suitable
technique is always required. Apart from this, there is no any forecasting methodology to use in
the company. With the help of this project, it will give a strong support for the planning.
 P a g e | 15

Chapter 3: Data Analysis & Interpretation

Brief introduction to time series forecasting models

Time Series is a collection of data points collected at constant time intervals. These are analyzed
to determine the long term trend so as to forecast the future or perform some other form of
analysis.
But what makes a TS different from say a regular regression problem? There are 2 things:

1. It is time dependent. So the basic assumption of a linear regression model that the
observations are independent doesn’t hold in this case.
2. Along with an increasing or decreasing trend, most TS have some form of seasonality
trends, i.e. variations specific to a particular time frame. For example, if you see the sales
of a woolen jacket over time, you will invariably find higher sales in winter seasons.

Because of the inherent properties of a TS, there are various steps involved in analysing it.

There are several techniques for forecasting from time series data. The following are three
important methodologies for time series forecasting. Other techniques are having some variations
over these models.
Moving Average (MA):
A moving average is a technique to get an overall idea of the trends in a data set; it is an average
of any subset of numbers. The moving average is extremely useful for forecasting long-term
trends. You can calculate it for any period of time. For example, if you have sales data for a
twenty-year period, you can calculate a five-year moving average, a four-year moving average, a
three-year moving average and so on. Stock market analysts will often use a 50 or 200 day
moving average to help them see trends in the stock market and (hopefully) forecast where the
stocks are headed.
An average represents the “middling” value of a set of numbers. The moving average is exactly
the same, but the average is calculated several times for several subsets of data. For example, if
you want a two-year moving average for a data set from 2000, 2001, 2002 and 2003 you would
find averages for the subsets 2000/2001, 2001/2002 and 2002/2003. Moving averages are usually
plotted and are best visualized.
Auto Regressive Integrated Moving Average (ARIMA):
ARIMA, short for ‘Auto Regressive Integrated Moving Average’ is actually a class of models
that ‘explains’ a given time series based on its own past values, that is, its own lags and the
lagged forecast errors, so that equation can be used to forecast future values.
Any ‘non-seasonal’ time series that exhibits patterns and is not a random white noise can be
modeled with ARIMA models.
 P a g e | 16

Holt-Winter Method:
Holt extended simple exponential smoothing to allow forecasting of data with a trend. It is
nothing more than exponential smoothing applied to both level (the average value in the series)
and trend. To express this in mathematical notation we now need three equations: one for level,
one for the trend and one to combine the level and trend to get the expected forecast ŷ

Using Holt’s winter method will be the best option among the rest of the models because of the
seasonality factor. The Holt-Winters seasonal method comprises the forecast equation and three
smoothing equations — one for the level ℓt, one for trend bt and one for the seasonal component
denoted by st, with smoothing parameters α, β and γ.

Data Analysis and Modeling

Data collected for most selling product HRCF for the last 3 FY 2017-2019. Data collected from
data ware house. The data preparation and cleaning tasks completed with due consideration. The
data collected is day wise and converted into monthly and quarterly data as required.
Key approaches for modeling:
i. Three models are tested i.e. Moving Average, ARIMA and Holt Winter’s Model
ii. The forecasting are made for selected products i.e. HRCF and GALCF based on most
sellable products.
iii. The model performance evaluated based on MAD (Mean Absolute Deviation).
iv. Moving Average tested on quarterly and half yearly basis.
 P a g e | 17

v. Prediction is done on monthly basis

Results are as below:

Method A. Moving Average
Product: HRCF

Product: GALCF

Method B: Holt Winter’s Method

Product: HRCF
 P a g e | 18

HW - Actual vs. Predicted

350000 120.00%

300000 100.00%

250000
80.00%
200000
60.00%
150000
40.00%
100000

50000 20.00%

0 0.00%

Abs_err pred NET_WT_MONTH

Product: GALCF

HW Actual vs Forecasted
50000 120.00%
45000
100.00%
40000
35000
80.00%
30000
25000 60.00%
20000
40.00%
15000
10000
20.00%
5000
0 0.00%

Abs_err pred NET_WT_MONTH

 P a g e | 19

Method C: ARIMA
Product: HRCF

ARIMA - Actual vs. Predicted

600000 160%
140%
500000
120%
400000
100%
300000 80%
60%
200000
40%
100000
20%
0 0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29

RESIDUALS PREDICTED ACTUALS

Product: GALCF

ARIMA (monthly level) - Actual vs. Preducted

45000 200%
40000 180%
35000 160%
30000 140%
120%
25000
100%
20000
80%
15000 60%
10000 40%
5000 20%
0 0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

RESIDUALS PREDICTED ACTUALS

 P a g e | 20

Data Interpretation
Model
Tests Level Sales Data Used Model Outline Outcome
Moving Month Order Booking - The current method being used to forecast sales This method when used on our data
Average Level Data since, we expect for each product. The method does help produces an average of 35%
Model - 4 this to represent the smoothen shocks/noise by using past sales. The between the forecasted & the actual
months true to sales level lag period used was 1 quarter values for HRCF & 54% for GAL. In
expected for each terms of the overall fit of the model,
month. the 6 months lag model beats the 4
months lag model
Moving Month Order Booking - The current method being used to forecast sales This method when used on our data
Average Level Data since, we expect for each product. The method does help produces an average of 40%
Model - 6 this to represent the smoothen shocks/noise by using past sales. The between the forecasted & the actual
months true to sales level lag period used was 6 months. values & 52% for GAL.In terms of the
expected for each overall fit of the model, the 6 months
month. lag model beats the 4 months lag
model
Holt-Winter's Month Order Booking - The results of the moving average indicate that The average % difference between
Model Level Data since, we expect more advanced method is required to forecast or the Actual & Forecasted is 22% for
this to represent the estimate the sales level. Shocks/Outliers from HRCF & 25% for GAL. The
true to sales level our data suggest that their impact needs to be comparison between the error from
expected for each captured as post a dip/spike, the sales pattern Holt-Winter's model & the moving
month. tends to decrease/increase. This in term implies average model, clearly shows that
a general trend where customers (across the HW model out performs the MA
industry segments) reduce/increase their by a considerable amount
consumption post a spike/dip.

The Holt-Winter's model uses smoothing

parameters & which enable the model the
control the impacts of previous months/data
points.
ARIMA Month Order Booking - The results of the moving average indicate that The average % difference between
Level Data since, we expect more advanced method is required to forecast or the Actual & Forecasted is 33% for
this to represent the estimate the sales level. Shocks/Outliers from HRCF & 40% for GAL. The
true to sales level our data suggest that their impact needs to be comparison between the error from
expected for each captured as post a dip/spike, the sales pattern Holt-Winter's model & the moving
month. tends to decrease/increase. This in term implies average model, clearly shows that
 P a g e | 21

a general trend where customers (across the HW model out performs the MA
industry segments) reduce/increase their by a considerable amount
consumption post a spike/dip.

The ARIMA uses an integration of moving

average & auto regressive method to forecast
the sales. This regression technique is used
when the today's data point is correlated to the
previous data point. It also like allows us to
capture impact of any seasonal changes
ARIMA Weekly Order Booking - The results of the moving average indicate that The actual vs. Predicted looks
Level Data since, we expect more advanced method is required to forecast or distorted. The model is unable to
(Pre & Post this to represent the estimate the sales level. Shocks/Outliers from explain any spike/dip. It usually over
Sunday roll true to sales level our data suggest that their impact needs to be predicts or under predicts the data
up) expected for each captured as post a dip/spike, the sales pattern
month. tends to decrease/increase. This in term implies
a general trend where customers (across
industry segments) reduce/increase their
consumption post a spike/dip.

The ARIMA uses an integration of moving

average & auto regressive method to forecast
the sales. This regression technique is used
when the today's data point is correlated to the
previous data point. It also like allows us to
capture impact of any seasonal changes. One of
the common requirements for ARIMA is that it
requires more number of data points. The thumb
rule is that it needs at least 30-35 data points.
The weekly model was tested to ensure that we
cover this point. In a weekly model, the impact of
a weekly shock/outlier is more difficult.
ARIMA Daily Level Order Booking - The results of the moving average indicate that The actual vs. Predicted looks
Data since, we expect more advanced method is required to forecast or distorted. The model is unable to
this to represent the estimate the sales level. Shocks/Outliers from explain any spike/dip. It usually over
true to sales level our data suggest that their impact needs to be predicts or under predicts the data
expected for each captured as post a dip/spike, the sales pattern
month. tends to decrease/increase. This in term implies
a general trend where customers (across
industry segments) reduce/increase their
 P a g e | 22

consumption post a spike/dip.

The ARIMA uses an integration of moving

average & auto regressive method to forecast
the sales. This regression technique is used
when the today's data point is correlated to the
previous data point. It also like allows us to
capture impact of any seasonal changes
Holt-Winter's Weekly Order Booking - The results of the moving average indicate that The smoothing parameters allow us
Model Level Data since, we expect more advanced method is required to forecast or to control the impact of the weekly
this to represent the estimate the sales level. Shocks/Outliers from shocks/outliers.
true to sales level our data suggest that their impact needs to be
expected for each captured as post a dip/spike, the sales pattern The % difference between the actual
month. tends to decrease/increase. This in term implies & predicted is around 9% for HRCF,
a general trend where customers (across while it is around 19% for GAL.
industry segments) reduce/increase their
consumption post a spike/dip. For GAL we recommend the monthly
HW model and not the weekly one.
The Holt-Winter's model uses smoothing This is due to the fact that the GAL
parameters & which enable the model the product exhibits significant dips
control the impacts of previous months/data which are not being explained by the
points. weekly level model. At a monthly
level they get evened out
 P a g e | 23

Chapter 4: Findings, Recommendations and Conclusions

Findings
1. Moving Average helps in smoothening the dataset but the accuracy is poor than other two
techniques. Thus when there are factors of seasonal trend this model will be less
effective.
2. Holt-Winter Models is used helps in smoothening the uncertain fluctuations of the
previous months thus able to improve the accuracy.
3. ARIMA model is though better than the MA but accuracy in comparison to the Holt-
Winter model is less.
Accuracy Matrix:

PRODUCT HRCF GALCF

MATRIX => MAE MAPE RMSE MAE MAPE RMSE
MODEL Value Value
Quarterly MA 47442 27.6% 55561 9688 49.9% 11122
Half-yearly MA 44708 25.2% 53364 8219 41.2% 9690
HW Model 30951 22.6% 44689 3599 19.6% 22440
ARIMA 59538 32.5% 83172 6877 40.2% 8881

Note:

Evaluation Matrix
MAE Mean Absolute Error
RMSE Root Mean Squared Error
MAPE Mean Absolute Percentage Error
Models
MA Moving Average
HW Holt Winter's Model
ARIMA Autoregressive Integrated Moving Average

Conclusions
Holt-Winter model is most suitable method for time series forecasting when there are dip spikes
in dataset along with the seasonal trend.
 P a g e | 24

References and Bibliography

References
Bibliography

Andersson, Rebecca, Jensen, Christian. (2017). How & Why Social Media is Used in B2B
Marketing. Linnaeus University.
Mari, A. (2018). The Rise of Machine Learning in Marketing: Goal, Process, and Benefit of
AIDriven. research.net, 3-7,19-20.
Marija Markovska, Aneta Buchkovska and Dimitar Taskovski. (2018). Comparative study of
ARIMA and Holt-Winters statistical models for prediction of energy consumption. Ss
Cyril and Methodius University, Faculty of Electrical Engineering and Information
Technologies. Skopje: researchgate.net. Retrieved from
https://www.researchgate.net/publication/326032579_Comparative_study_of_ARIMA_a
nd_Holt-Winters_statistical_models_for_prediction_of_energy_consumption
Moravick, A. (2018). UNDERSTANDING HOW MACHINE LEARNING WILL DRIVE VALUES
IN SALES. Aberdeen Group.
Nordlander, T. E. (2010). AI Surveying: Artificial Intelligence in Business. researchgate.net,
9,11,21,30.
people.ai. (2019). ACCELERATING SALES AND MARKETING EFFORTS THROUGH
ARTIFICIAL INTELLIGENCE. Harvard Business Review.