This document discusses predicting sales in supermarkets using machine learning techniques. It begins with an abstract that outlines the goal of developing an intelligent forecasting system using different machine learning techniques to build and adjust a sales forecasting model. The document then provides an introduction that discusses the importance of demand forecasting and challenges in predicting sales due to various internal and external factors. It reviews different forecasting techniques including time series models, exponential smoothing, neural networks and discusses using these approaches to more accurately predict future sales values.

1. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 03 | Mar-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 994
Predicting Sales In Supermarkets
Manjuprasad Shetty N1, Vishwasa Nawada N2, Kirthan Pai K3 , Ranjan Kumar H S4
1,2,3NMAM Institute of technology, NITTE
4 Asst. Professor, Dept. of Computer Science & Engineering, NMAM Institute of technology, NITTE,
Karnataka, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Most companies like to estimate the upcoming
trades. A superior forecasting can avoid them from over-
estimating or under-estimating the future trades which
leads to a great harm to the companies. Using reliable trade
calculation, companies may possibly assign their properties
more sensibly and make improved revenues. But projecting
the sales is rather complicated due to several internal and
external factors from the neighbouring atmosphere. Hence
in the current marketplace, there is an urgent requirement
for the progress of an smart forecasting system which is fast,
flexible and can provide high accuracy. We aim to put on
different machine learning techniques to construct and
adjust a sales forecasting model and perform estimation on
sales data to come across this requirement. We also intend
to provide a simple to use result with various visualization
tools for the easiness of users.
Key Words: Forecasting, Marketplace, Modelling
Trades ,Visualization.
1. INTRODUCTION
Request and deals anticipating is a standout amongst
other essential utilities of makers, suppliers, and
exchanging associations. Keeping interest and supply in
adjusting, they lessen abundance and deficiency of
inventories and enhance benefit. At the point when the
maker expects to satisfy the overestimated request,
overabundance generation brings about additional stock
keeping which ties up abundance stock. On the extra side,
underrated ask for causes unfulfilled requests, lost deals
foreordained shots and lessen benefit levels. The two
situations prompt wasteful store network. Therefore, the
exact request forecast is a genuine test for a challenger in
inventory network. Requests are needs for exact items
sponsored by a capacity to wage. Firms must ascertain
how various people require their antique as well as what
number of may really be sharp and fit for buying it.
Anticipating the ability to envision what customers will
perform thinking about the given arrangement of
conditions. Determining if the procedure of estimation in
obscure circumstances. A Forecast is a comparable,
however more broad term, and more often than not
alludes to estimation of time arrangement, cross-sectional
or longitudinal information. As of late, Estimating has
advanced into the act of Demand Scheduling in day by day
exchange anticipating for exchange firms. The teaching of
interest booking, otherwise called production network
prognosticating, holds both numerical anticipating and
exchange expectation isn't just about maintaining a gainful
adjust in-store network yet additionally to conjecture the
future deals in view of the styles and regular varieties. The
capacity of foreseeing the prospective fixated on past
information is an essential apparatus for support
individual and administrative basic leadership. In
particular, the objective of Time Series Forecasting (TSF)
is to device the execution of complex structures by taking
a glimpse at past outlines of a similar wonder.
Conventional estimating strategies experience the ill
effects of genuine constraints which trouble the
anticipating exactness. To modernize the conventional
ways a lot of new strategies and practices are being
connected by current information examiners.
2. LITERATURE SURVEY
There are two noteworthy types of forecasting. Macro
Forecasting and Micro Forecasting.
Macro Forecasting is unstable with foreseeing commercial
centres in entirety. This will be identified with close the
present level of Market Plea and assessing the results of
market request in the up and coming ages. Micro
Forecasting is concerned with entire segment deals
figures. This will be tied in with moulding a thing's
business sector parcel specifically industry and thinking
about the outcomes of piece of the overall industry later
on period. A gauge is an evaluation of an episode which
will occur in future. The forecast esteem isn't a
deterministic measure. Since it is just a gauge in light of
the past information identified with a specific occasion,
legitimate care must be given in evaluating it.
All the practical chiefs in any association will construct
their choices with respect to the figure esteem. Along
these lines, it is fundamental data for the affiliation.
Because of these reasons, appropriate upkeep ought to be
executed while evaluating estimate models.
2.1 Classification of forecast in business
In business, forecasts may be classified into technology
forecast, economic forecasts, and demand forecasts.
2.1.1 Technology Forecast
Technology is a mixture of hardware and software.
Hardware consists of corporal artefact while software is
the knowledge, practice or process. Technology forecast

2. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 03 | Mar-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 995
comes up with certain individualities such as level of
procedural act, rate of technical improvements.
Technological forecast is an estimation of the upcoming
features of useful mechanisms, products, method, events
or practices. Based on the reputation of this action, The
Government of India has created a “technology
information forecasting and assessment council (TIFAC)”,
under the Ministry of Science and Technology to
encourage action depended research and forecasting in
selected areas.
2.1.2 Economic Forecasts
Government assistance and other administrations enrol in
gathering data and forecasting of estimate on the general
commercial atmosphere. This will be convenient to
government organizations in foreseeing future tax
revenues, level of business growth, level of occupation,
level of rise, etc. Adding to this, there will be advantageous
to corporate groups to plot their upcoming movements
based on the level of commercial progress.
2.1.3 Demand Forecast
The Demand forecast provides the estimated level of
claim for merchandises or amenities. This is the simple
input for business planning and controlling. Hence, the
decisions for all the functions of any business firm are
affected by the demand forecast.
Demand forecasting has been pointed as an significant and
a stimulating problem for supply chain management [1].
This has captivated the consideration of numerous
investigation works. Past investigations have been
founded on the forecast of client request in view of time
arrangement models. A case for such are moving-normal,
exponential smoothing, and the Box-Jenkins strategy, and
easygoing models, for example, relapse and econometric
models.
Time Series Models can be isolated into two classes.
Univariate Models where the perceptions are those of
single variable recorded consecutively finished similarly
divided time interims. The other kind is the Multivariate,
where the perceptions are of different factors.
A typical approach for displaying univariate time
arrangement is the Autoregressive (AR) show. An
Autoregressive model is simply a direct relapse of the
present estimation of the arrangements in logical
inconsistency of at least one past estimations of the
arrangement. AR models can be investigated with one of
numerous approaches; including standard linear least
squares techniques. They also contain a forthright
understanding.
Models another basic approach for demonstrating
univariate time arrangement models is the moving normal
(MA) show. Moving normal model is reasonably a direct
relapse of the present estimation of the arrangement
against the repetitive sound irregular stuns of at least one
earlier estimations of the arrangement. The arbitrary
stuns at each point are accepted to originate from a similar
appropriation, commonly a typical circulation, with area at
zero and consistent scale.
Exponential Smoothing Method is a mainstream plan to
generate a smoothed Time Series. Though in Single
Moving Averages the past insights are weighted similarly,
Exponential Smoothing doles out exponentially
diminishing weights as the observation get more
seasoned. At the end of the day, late perceptions are given
moderately more weight in determining that the more
established perceptions.
Customary factual methodologies, including moving
normal, exponential smoothing, and autoregressive
incorporated moving normal (ARIMA) are straight in that
expectations without bounds esteems are obliged to be
direct elements of past perceptions. Second class is
nonlinear systems. In issue of time arrangement
estimating, the method of auto backwards coordinated
moving normal or ARIMA show broadly has been utilized
for foreseeing future esteem [2, 3].
Neural Networks (NNs) are adaptable non-straight
information determined models that have appealing
properties for determining. Factual strategies are
productive for information containing regular or pattern
designs, while counterfeit neural systems can suit the
information affected by the extraordinary case, similar to
advancement or outrageous emergency request
vacillation. Simulated Neural Network (ANN) [4, 5] can
learn like people, by collecting information through dreary
learning exercises. Creature mind's intellectual learning
process is reenacted in ANNs.
In the wake of investigating the above existing works,
there are different methodologies like Artificial Neural
system, Game hypothesis, grouping and so forth proposed
by different eminent information investigators and
researchers. We expect to centre around the current time
arrangement methodologies and its varieties as our
business estimate comes about rely upon patterns and
regular changes.
Time arrangement gauging can be extensively isolated into
3 classes Statistical techniques, AI strategies, Hybrid
strategies. Factual strategies and AI techniques are less
adaptable and give attractive estimating. In this way, in the
current year experts utilize half breed techniques. In the
crossbreed strategies utilized today even now the
exactness is palatable. The Clients require an accurate
solution with appropriate results and better prediction. To
build an appropriate and accurate solution for our
supermarket sales forecast we are using an extended
version of hybrid model. We will use our own set of
techniques and combination of relevant techniques to
form a new model.

3. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 03 | Mar-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 996
3. FEW TERMS IN SALES FORECASTING
3.1 Sales Demand: Sales demand for an item or package
is the projected total trades capacity in a business for a
precise interval time in a well-defined location, under a
clear advertising system or consumption. It is a term
related with different stages of industry publicizing costs.
3.2 Sales Prediction: Sales Prediction is the anticipated
market request a particular level of industry promoting
expenses.
3.3 Sales Probable: Sales Probable is the determined
market request, ensuing from a very superior level of
business publicizing spending, where additionally
increments in spending can have little impact on demand.
3.4 Company Request: Company request is the
enterprise’s projected segment of market request for an
invention or a package at different stages of the business
publicizing effort in a specific time era.
3.5 Sales Potential: Sales potential is the expected
business trades of a item or package, based on cut of
market latent estimated by the firm.
3.6 Sales Prediction: Sales Prediction is assessed
business sales of a service or merchandise, based on a
preferred publicizing costs plan, for an exact time interval,
in an anticipated marketing atmosphere.
3.7 Sales Expenses: Sales Expenses is the assessment of
the probable sales work in divisions or incomes from the
firm’s merchandises and facilities, and the retailing costs.
It is fixed slightly lesser than the firm sales prediction, for
avoiding unnecessary risks.
4. PROBLEM FORMULATION AND THEORETICAL
MODELING
Sales forecast is not just about maintaining a profitable
balance in supply chain but also to predict the future sales
based on the trends and seasonal variations. So, we are
proposing a new model which is a combination of various
time series approaches.
4.1 Time Series: Moving Normal Model
The Moving Normal model uses the last t time frames with
a specific end goal to anticipate request in period t+1.
There are two kinds of moving normal models:
straightforward moving normal and weighted moving
normal. The moving normal model presumption is that the
most precise expectation of forthcoming request is a
straightforward (direct) blend of previous request.
In the simple moving normal models, the projection value
is
T is the present time frame.
Ft+1 is the conjecture for following time frame.
n is the anticipating limit
A is the real transactions figure from each period
4.2 Time Series: Weighted Moving Average Model
t is the present time frame.
Ft+1 is the conjecture for following time frame
n is the anticipating horizon
At-n is the real deals figure from every time frame
w is the significance we provide for every time frames
4.3 Time Series: Exponential Smoothing (ES)
Model
The expectation without bounds depends for the most
part on the latest perception, and on the mistake for the
most recent gauge. Smoothing consistent alpha (α) means
the significance of the past blunder.
Considering that we are in period t. We ascertained the
conjecture for the last time frame (Ft-1) and we know the
real request last period (At-1)
4.3.1 Exponential Smoothing With Trend Model
FIT: Forecast Including Trend
δ: Trend smoothing constant
5. REAL WORLD CASE STUDY OF THE
SUPERMARKET SAMPLE DATA
We have collected data from Daily Needs Supermarket as a
source of data for our analysis. At this point of time, we
have divided out actual data into a sample and applied the
theoretical formulas that we will be using in the proposed

4. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 03 | Mar-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 997
model. We have also generated graphical representations
of the results for the above-mentioned samples.
Table-1: Actual Data showing sales of milk packets per
month
Above are the actual values we obtained from the data we
have collected. We calculated the forecasted result without
and with the trend using the above-mentioned formulas.
The results are tabulated in the below table.
Table -2: Actual and Forecasting results with and without
trend analysis trend analysis
The line graph below represents the above table in
graphical form which helps us to easily understand the
variance between the real results and the forecasted
outcomes.
Chart-1: Line Graph Representing Actual and Forecasted
number of Milk Packets sold vs. Month
6. CONCLUSION
From the results we obtained from the sample data, we can
conclude that our proposed model gives the supermarket
an accurate forecast about the sales. Even though the
results are not exact it is more than sufficient to help in
improvising supermarkets expansion strategies. Since the
forecasted results are so close to the actual results it also
helps in maintaining the supply chain balance. This model
can further be improvised by working on the accuracy of
the forecasted results using the hybrid model of Statistical
and AI methods.
REFERENCES
[1] İrem İşlek and Şule Gündüz Öğüdücü, “A Retail
Demand Forecasting Model Based on Data Mining
Techniques”, IEEE 24th International Symposium on
Industrial Electronics (ISIE), 3-5 June 2015.
[2] Soheila Mehrmolaei and Mohammad Reza
Keyvanpour, “Time series forecasting using improved
ARIMA”, Artificial Intelligence and Robotics
(IRANOPEN), 9th April 2016.
[3] Tsan-Ming Choi, Chi-Leung Hui and Yong Yu,
“Intelligent Time Series Fast Forecasting for Fashion
Sales: A Research Agenda”, International Conference
on Machine Learning and Cybernetics (ICMLC), 10-13
July 2011.
[4] Ilham SLIMANI, Ilhame El Farissi, and Said ACHCHAB,
“Artificial Neural Networks/or Demand Forecasting:
Application Using Moroccan Supermarket Data”, 15th
International Conference on Intelligent Systems
Design and Applications (ISDA), 14-16 Dec. 2015.
[5] Prasanna Kumar, Dr Mervin Herbert, and Dr Srikanth
Rao, “Demand Forecasting Using Artificial Neural
Network Based on Different Learning Methods:
Comparative Analysis”, International Journal For
Research In Applied Science And Engineering
Technology (IJRASET), Vol. 2 Issue IV, April 2014.
Month Milk Packets
January 1,324
February 1,352
March 1,304
April 1,276
May 1,209
At Ft Tt FITt
January 1324 1382 -11 1371
February 1353 1335 -26 1309
March 1304 1346 -7 1339
April 1276 1311 -21 1290
May 1209 1279 -27 1252