Intelligent Sales Prediction Using Machine Learning

Techniques
Sunitha Cheriyan Shaniba Ibrahim Saju Mohanan Susan Treesa
IT Department IT Department IT Department IT Department
Higher College of Technology Higher College of Technology Higher College of Technology Higher College of Technology
Muscat, Sultanate of Oman Muscat, Sultanate of Oman Muscat, Sultanate of Oman Muscat, Sultanate of Oman
[email protected] [email protected] [email protected] [email protected]

Abstract— Intelligent Decision Analytical System requires effectively, predictive sales data is important for businesses
integration of decision analysis and predictions. Most of the when looking for acquiring investment capital. The studies
business organizations heavily depend on a knowledge base proceed with a new perspective that focuses on how to
and demand prediction of sales trends. The accuracy in sales choose an appropriate approach to forecast sales with high
forecast provides a big impact in business. Data mining degree of precision. Initial dataset considered in this
techniques are very effective tools in extracting hidden research had a large number of entries, but the final dataset
knowledge from an enormous dataset to enhance accuracy and used for analysis having much smaller size compared to the
efficiency of forecasting. The detailed study and analysis of original due to the riddance of non-usable data, redundant
comprehensible predictive models to improve future sales
entries and irrelevant sales data.
predictions are carried out in this research. Traditional
forecast systems are difficult to deal with the big data and The data mining techniques and predictions methods are
accuracy of sales forecasting. These issues could be overcome discussed in Section I. The review of various literatures
by using various data mining techniques. In this paper, we about sales forecasts are stated in Section II. In Section III,
briefly analyzed the concept of sales data and sales forecast.
data tuning process and predictions are highlighted with
The various techniques and measures for sales predictions are
described in the later part of the research work. On the basis visual representation of generated results. The predictive
of a performance evaluation, a best suited predictive model is analytics and methodology on sales price also discussed.
suggested for the sales trend forecast. The results are The performance evaluations of various prediction
summarized in terms of reliability and accuracy of efficient algorithms using machine learning approaches are stated.
techniques taken for prediction and forecasting. The studies Finally, the result is analyzed and concluded by
found that the best fit model is Gradient Boost Algorithm, summarizing the research findings and future scope.
which shows maximum accuracy in forecasting and future
sales prediction. II. RELATED WORK

Keywords— Data mining techniques, Machine Learning

In order to be competent enough and to generate higher
Algorithms, Prediction, Reliability, Sales forecasting revenue, business organizations are constantly in search of a
better model or technique for data mining and maintenance
of critical data [2]. Business industry faces severe challenges
I. INTRODUCTION to identify an accurate data mining technique and effective
One of the major objectives of this research work is to predication strategy [3] due to the exponential growth of
find out the reliable sales trend prediction mechanism which huge volume of data used in e-commerce transactions. Sales
is implemented by using data mining techniques to achieve data analysis faces lot of issues and major aspects of sales
the best possible revenue. Today’s business handles huge functions are identification of product attribute, price
repository of data. The volume of data is expected to grow fixation, net sales realization and launch of new product.
further in an exponential manner. The measures are Various prediction methods, sales forecasting strategies and
mandatory in order to accommodate process speed of Expectation Maximization (EM) algorithm are discussed in
transaction and to enhance the expected growth in data [4].
volume and customer behavior. The E-commerce industry is A comparative study on data tuning and various
badly in need of new data mining techniques and intelligent clustering algorithms on sales data is clearly explained in [5].
prediction model of sales trends with highest possible level As analyzed in [6], classification of data is very important in
of accuracy and reliability. Sales forecasting gives insight decision making. Clustering techniques are very useful in
into how a company should manage its workforce, cash flow discovering distribution patterns and clustering algorithms
and resources. It is an important prerequisite for enterprise employ a distance metric based similarity measures [7]. In an
planning and decision making. It allows companies to plan appropriate data mining techniques information from a bulky
their business strategies effectively. data set can be transformed into a reasonable format and can
Accurate predictions allow the organization to improve be done by using supervised and unsupervised learning [8].
market growth with higher level of revenue generation. Data With an appropriate sales prediction technique, effective
mining techniques are very effective in tuning huge volume business decision making can be done. The concepts and
of data into useful information for cost prediction and sales algorithms are handled in [9]. As suggested by Korolev and
forecast, it is the basic of sound budgeting [1]. At the Ruegg, the prediction error can be reduced with the
organizational level, forecasts of sales are essential inputs to implementation of XGBoost and additional support of
many decision making activities in various functional areas SigOpt Bayesian Optimization method [10]. Sales
such as operations, marketing, sales, production and finance. forecasting can be done using different datamining
In order to serve an organization’s internal resources techniques where predicting sales on any given day at any

978-1-5386-4904-6/18/$31.00 ©2018 IEEE 53

store can be carried out, the detailed analysis and procedures
are shown in [11].
In this project we have performed sales forecasting for
stores using different data mining techniques. The task
involved predicting the sales on any given day at any store,
in order to familiarize ourselves with the task we have
studied previously.

III. RESEARCH METHODOLOGY

The main purpose of this research is to evaluate and
analyze the use of data mining techniques for sales
forecasting, to produce models which are comprehensive and
reliable.

A. Data Collection and Preparation

The dataset used for this research is based on an e-
fashion store, for the three consecutive years of sales data.
To predict the sales of the e-fashion store, past sales record
for three years from 2015 to 2017 were collected. The
database includes Category, City, Type of items and its
description, number of items, Quantity, Quarter, Sales
Revenue, Year, SKU description, Week, Year. The data
consisted initially of large number of entries, but the final
selected dataset had much smaller size compared to the
original dataset due to removal of non-usable data, redundant Figure 2. Yearly Sales Visualization
entries and irrelevant data[12].
The Figure 2 shows the visualization about the amount
B. Exploratory Analysis of sales revenue generated in the years from 2015 to 2017
After data preprocessing, in order to clearly understand and in the quarters represented as 1,2,3,4 respectively. This
the nature of our data, an exploratory analysis was shows that in quarter 4 of 2016 the sales revenue is high and
conducted[13]. The exploratory analysis consists of the steps in quarter 3 of 2015 shows a great decrease in the revenue
as shown in the Figure 1. generated.
C. Outlier detection
This process performs all necessary data preprocessing
and model optimization. Outlier detection process can be
used to deploy the model or as a starting point for further
optimizations and helpful in showing generic information
which is independent of the models. The main focus is on
the quality of the data, especially the quality of each data
attributes. Besides, these also consider discarding the data
attributes that provide less value.
Data: the dataset after it has been transformed for
modeling.
Correlations: a matrix showing the correlations between
the attributes with a positive correlation on the sales revenue
which is given in the Table II.
Figure 1. Stages of Data mining
The stages involved in the data mining model include TABLE II. CORRELATION MATRIX
data understanding, preparation, modelling, evaluation and
Attribute Quantity Sales SKU Week
deployment. The analysis of the collected sales data is shown
Revenue Description
in the Table I.
Quantity 1 0.947 -0.294 -0.021
TABLE I. YEARLY SALES DATA Sales 0.947 1 -0.247 -0.044
Revenue
sales Year of Data
SKU -0.294 -0.247 1 -0.269
Quarter 2015 2016 2017
Description
1 2,425,084 3,338,276 3,032,690
2 1,623,062 2,852,142 2,447,278 Week -0.021 -0.044 -0.269 1
3 1,254,177 2,885,557 3,306,000
4 1,795,108 4,199,956 2,107,128

54
D. Forecasting and Trends
The forecast is composed of a smoothed averaged adjusted
The figure 3 shows the forecasting of the future sales from for a linear trend. Then the forecast is also adjusted for
Quarter 3 of 2018 to Quarter 3 of 2021. The trend shows the seasonality. The Figure 5 shows the details about the model
sum of sales revenue for the dated Quarter. The blue color used for the trend analysis. The model shows a seasonal
indicates the actual sales generated and red color indicates effect high in the month of January 2022 and low in the
the estimated sales for the dated quarter showing a slight month of August 2022.
increase in the sales as shown in Figure 3.

Figure 3. Forecast for five years

Further the trend is also generated by grouping the

Sales Revenue and the sum of Quantity sold for quarter into
Cluster 1 and Cluster 2 respectively. The blue color line
indicates Cluster 1 and the orange color represents the
Cluster 2. The result shows a trend line showing a slight dip
in the Quantity sold and the Sales revenue in the Cluster 1
and 2 in the fourth Quarter as shown in the Figure 4.

Figure 5. Trend Analysis using Exponential Smoothing

E. Prediction

Prediction deals with events occurring in the future.

The use of Machine learning algorithms improves the
intelligence of the system without manual intervention.
“Machine Learning (ML) is used to optimize the
performance criterion using sample data or the past
Figure 4. Trend Analysis using Clusters experience” as defined by Ethem Alpaydin [14].

55
 Machine learning techniques can be applied to all
disciplines. Machine learning uses statistics to solve many ----------- (3)
classification and clustering problems. The ML algorithms Generalized linear models are providing estimate of the
are classified in three categories [15]. They are supervised, regression coefficients and estimated asymptotic standard
unsupervised and semi supervised. In this paper we errors of the coefficients. Usually the dispersion parameter
discussed about three machine learning algorithms which in GLS is fixed to a numeric value 1[18].
can be applied to prediction, like Generalized Linear Model
(GLM), Decision Tree (DT) and Gradient Boost Tree 2) Decision Tree
(GBT).
Decision tree is a classifier referred as recursive
partition of the instant space. It is a powerful form of
multiple variable analyses and is a strong data mining tool.
Its applications are found in various domains and this
approach represents factors involved in achieving a
predetermined goal and the corresponding factors to achieve
the goal and the ways and means of implementation. [14]
Let the objective can be denoted as (O) and (Ci) is the ways
to follow and let (Mij) the means of action corresponding to
these ways, which can be noted by qi, (i= 1 …. n), which
meets the relation.
…………….. (4)

For the means of action (Mij), the important coefficient (aij),

includes set of weights, where the sum is equal to 1 for each
way a11+a12+…..+a1m = 1,
a21+a22+…..+a2m = 1
…..…..
an1+an2+……..+anm= 1

……………… (5)

The Figure 7 is an example of decision tree model

displaying the different items as a tree in each quarter and
the root node is the year.

Figure 6. System Architecture

In this study, we implemented three machine learning

algorithms on the training dataset and the models are tested
for the performance. Based on the performance accuracy
the best algorithm is chosen for the prediction.

1) Generalized Linear Model

Generalized linear model (GLM) refers to a large class of

conventional linear regression model [16]. The focus is for a
Figure 7. Decision Tree Model
continuous response where the variable gives continuous
categorical predictors [17]. One of the major components in
3) Gradient Boosted Trees
a generalized linear model is a random component which is
the probability distribution of the response variable (Yi);
Gradient boosting is a machine learning technique for
A linear predictor is another important component, which
regression and classification problem. This approach could
can be represented
p as:
ensemble learning method that combines large number of
----------- (1) decision trees to produce final prediction model [10]. This
A smooth and invertible linearizing link function g(.), which model is built on a principle that a collection of weak
transforms the learners combined together can produce a strong learner by
----------- (2) using boosting process. GBT approach has a strong additive

56
training method, required for adding a new weak learner further improvement on the GBT implementation with the
into the model, the weak learner is the decision tree [19]. support of a strong data set along with models such as
Grabit, Tobit as analyzed in [20], projects better accuracy
Let F(x) is a full model after t-1 round and h(x) is the new rate.
tree, added to the model. F0= 0 ………. (6)

Ft (x) = Ft-1(x) +h(x) ...……... (7)

Each new function is an attempt to correct the

errors of the model built in previous rounds. Hence the new
function(x) must be able to predict the residual Ft-1(x). The
Figure 8 shows a sample of gradient boosted tree.

Figure 9. Comparison between the Models

TABLE III. PREFORMANCE SUMMARY

Performance Summary of ML Algorithms

Model
Name Accuracy Error
Rate (%) Rate Precision Recall Kappa1
GLM 64 36 5.36 0 0

DT 71 29 11.24 15.61 0.501

GBT 98 2 50 50 0.962

The accuracy rate, error rate, precision, recall and Kappa

values shown in Table III is a clear indication that Gradient
Boosted Tree (GBT) stand out as pioneer model with the
Figure 8 . Gradient Boosted Tree Model
accuracy rate of 98 % and minimum error rate.

V. CONCLUSION
IV. RESULT AND ANALYSIS
The researchers have concluded that an intelligent sales
prediction system is required for business organizations to
The performance of the classification algorithms is
handle enormous volume of data. Business decisions are
mostly focused on Classification accuracy, Accuracy in each
based on speed and accuracy of data processing techniques.
class and confusion matrix which shows the number of
Machine learning approaches highlighted in this research
predictions of each class which can be compared to the
paper will be able to provide an effective mechanism in data
instances of each class. Root Mean Square Error, Mean
tuning and decision making. In order to be competent in
Square Error, Absolute error are calculated and average of
business, organizations are required to equip with modern
the error is shown in the output in the Table III as the Error
approaches to accommodate different types of customer
Rate. This measure helps to identity whether the given
behavior by forecasting attractive sales turn over. In our
prediction is wrong on average.
studies, we used almost 85,000 records for the comparison
of algorithms. Since the time of execution was huge and to
The comparative studies of the three algorithms based on
manage such a large set of records are complex, some of the
the prediction performance are given in the Table 1 and the
records were discarded, during the analysis phase. At the
visualization in Figure 9. Based on the performance, it is
same time, fields and attributes, used in this analysis were
understood that Gradient Boost Algorithm is showing 98%
insufficient for the further analysis. It was the major
overall accuracy and the second stands Decision Tree
challenge we faced during the research. However, we had
Algorithms with nearly 71% overall accuracy and followed
thoroughly weighed our works by implementing efficient
by Generalized Linear Model with 64% accuracy. Finally,
ML techniques for prediction and forecasting. The current
it can be compared based on the empirical evaluation of the
studies can be expedited by using Big Data as a tool for the
three chosen algorithm the best fit for the model is Gradient
predictive analytics in sales forecasting. The big data
Boosted Tree. The classification accuracy rate can reach up
analysis and forecasting are measured as the vital fields in
to 100%, but in GBT model analyzed and shown in Table
the modern business scenario.
III, achieved approximately 98% of accuracy. If there are

57
 REFERENCES [12] Rey, T. D., Wells, C., & Kauhl, J. (2013). Using data mining in
forecasting problems. In SAS Global Forum 2013: Data Mining and
Text Analytics.
[1] Huang, Q., & Zhou, F. (2017, March). Research on retailer data [13] Huang, W., Zhang, Q., Xu, W., Fu, H., Wang, M., & Liang, X.
clustering algorithm based on spark. In AIP Conference Proceedings (2015). A Novel Trigger Model for Sales Prediction with Data
(Vol. 1820, No. 1, p. 080022). AIP Publishing. Mining Techniques. Data Science Journal, 14.
[2] Saylı, A., Ozturk, I., & Ustunel, M. (2016). Brand loyalty analysis [14] Ethem Alpaydin. (2004). Introduction to Machine Learning (Adaptive
system using K-Means algorithm. Journal of Engineering Technology Computation and Machine Learning), The MIT Press.
and Applied Sciences, 1(3).
[15] Lytvynenko, T. I. (2016). Problem of data analysis and forecasting
[3] Maingi, M. N. A Survey on the Clustering Algorithms in Sales Data using decision trees method.
Mining.
[16] Lazăr, C., & Lazăr, M. (2015). Using the Method of Decision Trees in
[4] Sastry, S. H., Babu, P., & Prasada, M. S. (2013). Analysis & the Forecasting Activity. Petroleum-Gas University of Ploiesti
Prediction of Sales Data in SAP-ERP System using Clustering Bulletin, Technical Series, 67(1).
Algorithms. arXiv preprint arXiv:1312.2678.
[17] Flesch, B., Vatrapu, R., Mukkamala, R. R., & Hussain, A. (2015,
[5] Shrivastava, V., & Arya, N. (2012). A study of various clustering October). Social set visualizer: A set theoretical approach to big
algorithms on retail sales data. Int. J. Comput. Commun. Netw, 1(2). social data analytics of real-world events. In Big Data (Big Data),
[6] Rajagopal, D. (2011). Customer data clustering using data mining 2015 IEEE International Conference on(pp. 2418-2427). IEEE.
technique. arXiv preprint arXiv:1112.2663. [18] Asooja, K., Bordea, G., Vulcu, G., & Buitelaar, P. (2016). Forecasting
[7] Tsai, C. F., Wu, H. C., & Tsai, C. W. (2002). A new data clustering Emerging Trends from Scientific Literature. In LREC.
approach for data mining in large databases. In Parallel Architectures, [19] Stearns, B., Rangel, F., Rangel, F., de Faria, F. F., Oliveira, J., &
Algorithms and Networks, 2002. I-SPAN'02. Proceedings. Ramos, A. A. D. S. (2017). Scholar Performance Prediction using
International Symposium on (pp. 315-320). IEEE. Boosted Regression Trees Techniques. In European Symposium on
[8] Mann, A. K., & Kaur, N. (2013). Review paper on clustering Artificial Neural Networks, Computational Intelligence and Machine
techniques. Global Journal of Computer Science and Technology. Learning (ESANN). Citeseer.
[9] Shah, N., Solanki, M., Tambe, A., & Dhangar, D. Sales Prediction [20] Sigrist, F., & Hirnschall, C. (2018). Gradient Tree-Boosted Tobit
Using Effective Mining Techniques. Models for Default Prediction.
[10] Korolev, M., & Ruegg, K. (2015). Gradient Boosted Trees to Predict
Store Sales.
[11] Jain, A., Menon, M. N., & Chandra, S. Sales Forecasting for Retail
Chains.

58