CHAPTER ONE

INTRODUCTION

1.1 BACKGROUND OF THE STUDY

Outstanding business planning relies on how much appeal there will be for outputs and treatments that

can be recognized by leaders. The more insight into demand from product design to production,

logistics to sales, and the better prepared the organization will be. On the contrary, if the forecasts

were unsatisfied, it will result in the risk of over or under production, weakness in service, or a mere

selling of the faulty commodity. At least to achieve maximum income, market opportunities with the

greatest potential demand should be pursued. Apart from that getting the right quantity of product at

the right time at the right place (production), planning the right production volume and inventory

(inventory), and optimizing the supply chain for delivery (supply chain). Besides, finding valuable

prospects using long-term customer profiles to identify and get prospects with similar characteristics

is extremely important to do. This is where forecasting should begin to be carried out by an

organization. Valuable outlook will seize authentic patterns and correlations that be found in historical

data but not renew past events that will not occur again, whereas forecasting methods be able to be as

common as using last review s such as forecasting (called naive methods), or high complexity, such as

neural networks and econometric system of simultaneous equations ( Hyndman RJ and

Athanasopoulos, 2018). Having rigorous demand forecasting and stocktaking management systems

for vulnerable operational management is essential because of rivalry and financial constraints in the

retail industry, as well as cost-oriented supply chain operations, and retailers entail to optimize their

provision to degrade the financial risk ( Kilimci, 2019). Even though circumstances such as demand

uncertainty could be defying, a long-term plan might allow an opportunity for demand planning

supported by divination and arrangement that put aside time, turn downs costs, and sharpens

outcomes (Van derLaan, et at., 2016). Supply chain analysis is needed in a company because it can

help in the process of its development, which includes the process of selecting suppliers, planning

logistics, and distributing supplies. The existence of supply chain analysis will have an important

effect and have a positive impact on the company, such as getting satisfaction from consumers. If

1
customers are satisfied, then they become loyal customers who use your product for a long time.

Besides that, there is also another positive impact, namely increased profit. If there are more loyal

customers, the company's profits will also increase because the products offered are sold out.

Therefore, business people are expected to be proficient in analyzing data based on the purchase

history and profitability of each client, because the profitability is possible to build up over time

(Chen, Guo, and Ubakanma, 2015). Predicted time series is one of the most preferred but throw

down the most gauntlet and has played an important role in all kinds of fields including Sta tistics,

Machine Learning, Data Mining, and Operations Research for decades (Shi et al., 2020).

Customer satisfaction is a very vital parameter of business. Retailers always try to fulfill their

demands as often and possible while continually looking to earn more profit with a calculated

investment to increase sales. Therefore, product sales forecasting helps retailers to increase

profit with absolute investment.

Forecasting or prediction refers to the process of trying to foretell the likelihood of future

events from past and present data available by analysis of trends and patterns found within

the data (Krishna et al., 2018). In the same light, sales prediction is the process of estimating

future revenue for a firm or outlet by predicting the amount of revenue to be generated over

the given period in review (Majhi, 2019). This period may differ from weeks to months or

even years. For any business to be productive, it is expected that they make judicious use of

resources available to them to make sound strategic decisions for the future to remain

competitive and increase their revenue base (Universitesi, 2017).Since it is known that all

forecasts involve at least some level of uncertainty, businesses need to make estimates with

the aim of minimizing this uncertainty. Therefore Organizations should not make predictions

considering a few factors, but their predictions should comprise various variables such as

raw-material requirements, optimal stock levels, borrowing conditions, and personnel

conditions (Universitesi, 2017). Furthermore, in order for any of these to be projected, it is

crucial to predict the expected demand that will apply to the market and, appropriately, the

2
company's potential sales (Universitesi, 2017). Thus, accurate forecasts allow for appropriate

targeting of the company's activities (such as production, finance, R&D, purchasing, and

marketing) and assist them in reaching their targets (Mentzer, et al. 1998).

A significant number of business problems require decisions to be made based on a large

amount of historical data and human decision-makers have difficulty integrating these data;

thus, their decisions are susceptible to several biases (Jagielskaand Jacob, 1993). Machine

learning algorithms and data mining have been employed in recent times to help

organizations and businesses combat and tackle the problem of harvesting data and using the

knowledge found within them to make predictions for sales. With a good data mining

infrastructure in place and the right application of predictive algorithms, companies stand a

better chance of making better-informed decisions. Therefore, Sales forecasting is deemed of

paramount importance for companies with intentions of entering new markets or adding new

services, products, or experiencing high growth. When guided and assisted with the right

tools to perform the task of sales forecasting, organizations tend to exert control and portray

an outstanding performance in the market, thereby ensuring minimal loss of revenue and

increased profit (Jagielskaand Jacob, 1993). It is assuring to note that the main reason a

company does a forecast is to balance marketing resources and sales against supply capacity

planning (Jagielskaand Jacob, 1993). With the right implementation of forecasting,

companies and corporations no doubt can address fundamental questions such as "can we

drive demand with our current fusion of price, promotion, and marketing?", "are the

resources currently at our deposal adequate to measure up to demand?", "Is there enough

personal to match the volume of budgeted sales?" (Krishna et al., 2018). In this regard,

companies and organizations allocate and invest a reasonable amount of resources in both

human and finance to obtain adequate and genuine prediction results.

3
Data Mining and Machine Learning have been successfully devoted to the time series computation of

numerous amounts of demand-related data and features, and to forethought coming necessity and

prototype using particular learning theorem ( Kilimci, 2019). The challenges with big data are well

established and have spawned a whole new industry around cloud computing services that enable easy

access and relatively inexpensive computing solutions ( Bradlow, 2017). Demand preparation is more

about forecasting and expertise to expect the requirement for diverse goods at various points in the

stock line which allows an organization to better plan their inventory, ensure product availability

according to market needs, and monitor the difference between actual and predicted sales to optimize

their production.

1.2 STATEMENT OF THE PROBLEM

The existing forecasting systems encounter challenges in accurately predicting product

demand, leading to issues such as overstocking or stock-outs. These inaccuracies can result in

significant financial losses, decreased customer satisfaction, and operational inefficiencies.

Addressing these challenges is imperative for businesses striving to optimize their supply

chain operations.

1.3 AIM AND OBJECTIVES OF THE STUDY

This study's principal aim will be to implement data mining techniques and machine learning

algorithms to help identify the class of products that sell more for a business outlet to aid the

business owners in making more informed decisions on inventory acquisition. This would be

carried out by performing a comparative analysis of four different classification machine

learning algorithms in line with the CRIPS_DM methodology to see which algorithms best

perform on the available data. Data mining techniques and implementation shall be discussed

in the latter part of the research. This paper will consider a variety of forecasting algorithms

such as the Decision tree, deep learning, Naïve Bayes and Random Forest.

While the specific objectives are as follows;

4
  To investigate the current landscape of product forecasting in the business domain.

 To assess the limitations of traditional forecasting methods.

 To explore the principles and applications of deep learning in machine learning.

 To design and implement a machine learning-based product forecasting system.

 To evaluate the effectiveness of the developed system through empirical analysis.

1.4 SIGNIFICANCE OF THE STUDY

Sales forecasting adds excellent value to the business because it deals with the future. For

businesses in the private sector, the confidence level concerning their future investments is

made strong when the predictions are accurate and can be trusted (Cheriyan et al., 2019).

Similarly, for publicly traded companies, an accurate forecast will bestow strong credibility

in the market and adds value to the business (Cheriyan et al., 2019). Also, firms that are into

production employ forecasting to moderate and plan their production circle to ensure they

meet up with demand, curb excess and waste (Wang and Liu, 2019). In a scenario where

there is inaccurate forecasting or disconnect within the firm or business, and forecasts are not

communicated, the outcome is always adverse for the business as this will result in

discrepancies and gaps within the organizations' goals and targets. It is worthy to note that

when an accurate sales forecast is available for a business, it confers the following benefits.

• It guides the alignment of sales quotas and proceeds expectations.

• Helps to improve decision making about the future.

• It helps in reducing the time spent on planning territory coverage and establishing

quotas.

• It helps to set benchmarks that will be used to access trends in the future.

5
1.5 SCOPE OF THE STUDY

The following are the scopes the study will encompass as well as the various aspects related

to understanding, developing, implementing, and optimizing the system. Here are key

components within the scope of the study:

• Business Understanding:

Gain a deep understanding of the business objectives, product portfolio, market dynamics,

and key stakeholders' requirements. Identify the specific goals and use cases for the

forecasting system, such as inventory optimization, production planning, or revenue

forecasting.

• Data Collection and Preprocessing:

Identify relevant data sources, including historical sales data, product attributes, market

trends, and external factors. Clean, preprocess, and integrate data from disparate sources to

create a comprehensive dataset for analysis.

• Feature Engineering:

Extract meaningful features from the data that capture important patterns, trends, and

correlations relevant to product demand forecasting. This may involve time-based features,

product attributes, seasonal indicators, and lagged variables.

• Model Selection and Development:

Evaluate and select appropriate forecasting methods and algorithms based on the

characteristics of the data and the forecasting objectives. Consider traditional statistical

methods (e.g., ARIMA, exponential smoothing) as well as advanced machine learning

techniques (e.g., deep learning, ensemble methods).

• Model Training and Validation:

Train forecasting models using historical data and validate their performance using

appropriate evaluation metrics and validation techniques such as cross-validation or holdout

6
validation. Ensure that the models generalize well to unseen data and can provide reliable

predictions.

1.6 LIMITATION

The researcher may encounter several challenges when designing a product forecasting

system, which could impact the study and effectiveness of the system. Some of these

limitations include:

• Data Availability and Quality:

Limited historical sales data or incomplete datasets can hinder the accuracy of forecasting

models. Additionally, data quality issues such as missing values, outliers, and inaccuracies

may affect the reliability of predictions.

• Model Complexity and Interpretability:

Deep learning models and complex forecasting algorithms may require significant

computational resources and expertise to develop and maintain. Moreover, the interpretability

of such models may be limited, making it difficult to understand the underlying factors

driving the predictions.

• Assumptions and Limitations of Forecasting Methods:

Different forecasting methods have their own assumptions and limitations. For example, time

series methods may assume stationarity or independence of observations, which may not hold

true in real-world scenarios.

• Resource Constraints:

Constraints such as budget, time, and expertise can limit the scope and sophistication of the

forecasting system. Implementing and maintaining advanced forecasting techniques may

require investment in technology infrastructure and personnel training.

• Evaluation and Feedback Loop:

7
Establishing an effective evaluation framework to assess the performance of forecasting

models and incorporating feedback from actual sales data into the system is essential but can

be challenging due to delays in feedback or changing business conditions.

1.7 DEFINITION OF TERMS

• Forecasting: refers to the process of making predictions about future events or

conditions based on historical data and current trends.

• Prediction: refers to a statement or act of forecasting what will happen in the future

based on observations, experience, or scientific reasoning.

• Competitive: refers to a condition or characteristic that refers to, is based on, or is

characterized by competition.

• Business: is an organization or entity engaged in commercial, industrial, or

professional activities, involved in the production and sale of goods and services.

• Organizations: refers to a group of people who work together, often for a specific

purpose, such as a neighborhood association, a charity, a union, or a corporation.

• Sales: refers to activities related to selling goods or services, or the number of goods

within a specific time frame.

• Demand: refers to a firm request for something, whether it be goods, services, or

actions. It can also indicate the quantity of a commodity or service that is wanted at a

specific price and time.

• Product: is something that is produced and sold in large quantities, often as a result of

a manufacturing process.

• Targets: refers to a specific group of people with shared characteristics that a business

aims to reach with its products or services.

• Logistics: refers to the process of coordinating how goods and products are obtained,

stored, and distributed.

8
• Production: refers to the process of creating goods or services from raw materials,

components, or ideas.

• Customer: is an individual or business that purchases goods or services from another

company.

• Investment: refers to the outlay of money or capital with the expectation of earning

profitable returns, such as interest, income, or appreciation in value.

9
 CHAPTER TWO

LITERATURE REVIEW

2.0. INTRODUCTION

This section reviews the background concepts which includes; definitions of product

forecasting, Time series, Demand forecasting methods, inventory management, data analytics

and machine learning, performance evaluation, deep learning algorithms used for product

forecasting as well as previous works being done in this field.

2.1. REVIEW OF BACKGROUND CONCEPTS

According to (Hyndman et al., 2018), Forecasting is the process of making predictions or

estimates about future events based on past and present data. It involves analyzing historical

data patterns and trends to anticipate future outcomes.

Product forecasting refers to the prediction of future demand or sales for specific products or

product categories. It involves analyzing historical sales data, market trends, customer

behavior, and other relevant factors to estimate future demand levels for individual products

(Silver et al., 1998).

A product forecasting system involves various concepts such as time series analysis, demand

forecasting methods, and inventory management principles. Here are some key background

concepts:

• Time Series Analysis:

Time series analysis involves studying patterns and trends in data collected over time. It's

essential for understanding historical sales data and forecasting future demand (Jenkins,

Reinsel & Ljung, 2015).

10
 • Demand Forecasting Methods:

Demand forecasting methods include qualitative techniques (e.g., expert opinion),

quantitative methods (e.g., time series analysis, regression analysis), and causal methods

(e.g., econometric models) (Makridakis, Wheelwright and Hyndman, 1998).

• Inventory Management:

Inventory management principles involve optimizing stock levels to meet demand while

minimizing costs. This includes techniques like economic order quantity (EOQ) and just-in-

time (JIT) inventory management (Silver, Pyke and Peterson, 1998).

• Data Analytics and Machine Learning:

Data analytics and machine learning techniques are increasingly used for forecasting,

especially in large datasets where traditional methods may be less effective (James et al.,

2013).

• Performance Evaluation:

Performance evaluation measures assess the accuracy of forecasting models, such as mean

absolute percentage error (MAPE) and mean squared error (MSE) (Hyndman and Koehler,

2006).

These concepts form the foundation of a robust product forecasting system, helping

businesses make informed decisions about production, inventory levels, and resource

allocation.

2.1.1. DEEP LEARNING ALGORITHMS USED FOR PRODUCT FORECASTING

• Deep Learning:

11
Deep learning is a subset of machine learning that involves neural networks with multiple

layers, enabling the automatic extraction of features from data (LeCun, Bengio and Hinton,

2015).

• Recurrent Neural Networks (RNNs):

RNNs are neural network architectures designed to process sequential data by maintaining a

memory state across time steps (Lipton, Berkowitz and Elkan, 2015).

• Long Short-Term Memory (LSTM) Networks:

LSTMs are a type of RNN architecture that addresses the vanishing gradient problem by

maintaining a memory cell with gating mechanisms (Hochreiter and Schmidhuber, 1997).

• Gated Recurrent Units (GRUs):

GRUs are a simplified variant of LSTM networks that use gating mechanisms to control the

flow of information through the network (Cho et al, 2014).

• Sequence-to-Sequence Models:

Sequence-to-Sequence models are neural network architectures capable of mapping input

sequences to output sequences, commonly used for tasks like language translation and time

series forecasting (Sutskever, Vinyals and Le, 2014).

• Attention Mechanism:

Attention mechanisms enable neural networks to focus on relevant parts of the input

sequence when making predictions, improving performance in sequence modeling tasks

(Bahdanau, Cho and Bengio, 2014).

2.2. REVIEW OF RELATED WORKS

The search for better models and techniques for data mining by companies and organizations

is a continuous one as it aids them to remain competitive and generate higher revenue

(Cheriyan et al., 2019). There have been lots of work in the area of product sales forecasting

using machine learning and data mining. This section of the work goes through a brief survey

12
of several related studies that have been carried out in sales forecasting and related

forecasting problems. Many statistical models and techniques such as regression, (ARIMA)

Auto-Regressive Integrated Moving Average, (ARMA) Auto-Regressive Moving Average,

have been applied for creating multiple sales forecasting concepts. However, forecasting is a

complex problem and is influenced by external as well as internal factors, and there are

drawbacks to the statistical approach as told by Weigend et al. (1994). In their work, Krishna

et al. (2018) stated that "forecasting can be used to predict sales revenue at the product level,

or at an individual business level, or at a company level. However, their work concentrated

on product level sales forecasting. They explained that Future sales plan aids in optimal

utilization of facility, scheduling, conveyance, and effective control of inventory. These, in

turn, result in enhancement of clients' satisfaction and decrease in production cost. They

explained that sales forecasts affect a company's marketing plan directly and that the

marketing departments of businesses and companies are accountable for how their customers

understand its services or products and compare it against its competitors by using their sales

forecast to assess how marketing expenditure can increase sales and channel in more demand.

Thus, it is crucial to develop effective sales forecasting models in order to make accurate and

robust forecasting outcomes. However, they pointed out that in the business and economic

environment, it is imperative to correctly predict various kinds of economic variables like the

past economic performance, present global conditions, inflation rate, internal organizational

changes, marketing strategies, seasonal factors, etc. so as to develop appropriate approaches.

They considered a range of forecasting techniques such as multiple regression, Polynomial

regression, Ridge regression, and Lasso regression alongside a variety of boosting algorithms

like AdaBoost, Gradient tree boosting to get the maximum accuracy.

(Arif et al., 2019) presents a new method-using machine learning to assist with accurate

predictions in a comparative study. Their proposed method collects and analyzes the store's

13
previous data by implementing various algorithms like the K-Nearest Neighbor (KNN),

Gaussian Naïve Bayes, and Decision Tree Classifier to find out the best technique for

forecasting demand. Their analysis of these algorithms found out that Gaussian Nave Bayes

has the best accuracy for their historical data. In work done by S. Cheriyan et al. (2018), they

implemented various machine learning algorithms and data mining techniques with the

perception of choosing the best approach to forecast sales with a high degree of precision.

However, they pointed out that traditional forecast systems are complicated when dealing

with big data and the accuracy of sales forecasting. Nonetheless, they explained that these

issues could be surmounted by utilizing various data mining techniques.

Likewise, Chand N et al. (2017) in their paper, suggest a new set-up methodology using the

averaging approach, which not only gives priority to an algorithm that reliably retains good

precision but also reduces the variance from actual sales. Their results found that the

assembling of the time-series model results and the regression model gives a better result due

to the nullification of over-forecasting and under-forecasting and put the forecast values close

to the real in most situations and gives weightage to the model that has performed better

previously. Also in the research paper written by Lin K. Y et al. (2016), they developed a

deep learning-based customer forecasting tool to improve the management quality for

decision making. Their proposed approach extracted the relationship between the number of

customers and the leading factors of the vast amount of historical data. A Deep Learning

algorithm was used in executing sales analysis and produce steadfast predictions that would

become administrative tools, especially with the building of a forecasting model based on

customer behavior with weather as the prediction parameter. Their study built a deep

learning-based customer forecasting tool to analyze Point of Sale (POS) data. Their proposed

model allows business units to make the appropriate decisions that represent an improvement

in the store's performance and development that enhances the customers' experience.

14
Considering the lousy forecasting results of the standard support vector machine (SVM) for

product sale series with the normal distribution noise. (Wu Q et al., 2008) proposed an SVM

based on the Gaussian loss function. Then, a hybrid forecasting model for product sales and

its parameter-choosing algorithm is presented in their study. Wu, however, pointed out that

the standard SVM encounters several complexities in real-life application but new and

improved SVMs have been developed to help solve the actual problems of SVM.

Nonetheless, they further stated that the standard SVM adopting ε -insensitive loss function

has good generalization capability in some applications but challenging to deal with the

standard distribution noise parts of (time) series. Their work mainly focuses on establishing a

new SVM that can deal with the Gaussian noise parts of series.

In their study, Jagielska I et al. (1993) investigates the connectionist concept to sales

forecasting. They built a neural network model for the prediction of sales levels, and the sales

forecast produced was compared with the actual sales figures as well as with the estimates

made by the marketing managers. The model they built was presented with the past sales data

then the system trained on the data and learned the pattern from which it made its predictions.

In their study, they showed that neural networks could be applied to effectiveness in solving

problems that are complex and non-deterministic. The neural network they built was meant to

predict sales figures for Tattersall, which is Australia's leading gaming operator, to help their

marketing managers make sales forecast for lottery products.

Furthermore, the work carried out by Kilimci Z H et al. (2019) proposed the infusion of nine

different time series methods, Support vector regression algorithm (SVR), and Deep Learning

forecasting models to improve the demand forecasting process, which is one of the major

issues concerning supply chain. To achieve a final decision of these models for their

proposed system, they blended the algorithms with a novel integration strategy that is

reminiscent of boosting ensemble strategy. Also, they introduced another novel approach in

15
their work by adopting a boosting strategy to their demand forecasting model. In this way,

their proposed system's final decision was based on the best algorithms with regards to

gaining more weight. To this effect, their forecast was more reliable with respect to the

changes in trends and seasonality behaviors.

A sentimental analysis approach was used by Avinash K.S et al. (2020) in their work to

forecast the demand for a product for a particular period. Their study explained in detail how

random forest and clustering algorithms with neural networks could be used to implement

product demand forecasting to improve organizations or business profits.

Regarding the work done by Mortensen S et al. (2019), they develop a model that can

forecast sales performance with a fair degree of accuracy. They experimented with a range of

machine learning techniques, including binomial logic and various decision tree methods,

including gradient and random forest boosting. In another study carried out by Saena W et al.,

(2018) they applied machine learning methods for predicting the number of pharmaceutical

sales. The efficacy of each model was compared using their errors. CNN-LSTM model

produces better forecasting results for their study. With inventory planning being one of the

fundamentals of a retail fashion operation, an adequate forecasting system can undoubtedly

go a long way to help businesses in fashion retail avoid understocking or overstocking in

inventory planning. This further relates to other essential operations of the whole supply

chain like production planning, pricing and achieving customer satisfaction.

(Na Liu et al., 2013) presents a comprehensive review of the literature on fashion retail sales

forecasting and equally explores the benefits and the downsides of various analytical

techniques for fashion retail sales forecasting as well as examining the pertinent issues

relating to real-world implementation of fashion retail sales forecasting algorithms. In another

study, Dokuz Eylul universities et al. (2017) investigates the forecasting performances of

models such as state-space models and the ANFIS model, making them be the first to

16
contribute to the literature on the study. Their study analyses forecasts of the sales volume of

a retail furniture store in order to appraise the forecasting performance of five methods:

namely, state-space models, the ARIMA model, the ARFIMA model, the ANN model, and

the ANFIS model. Based on the accuracy observed for each of the forecasting techniques,

they suggested that employees in various relevant departments can derive a general

understanding of which forecasting approach conform best to which product sales and

thereby identify opportunities to ameliorate the existing supply chain system. They aimed to

determine which of the most employed blended techniques produce greater statistical

enhancements in the relevant series’ forecast accuracy. Their study concluded that it is

needless to identify the best for individual models when seeking to improve the forecast

accuracy achieved by combined forecasts. Furthermore, they claimed that a combination of

very similar models affects the performance of forecasting positively.

(Arif M A I et al., 2020) took a new machine learning approach that helps produce accurate

predictions. Their work focuses on finding the best algorithm to estimate product demand

with high precision and less error. They analyzed and used three popular algorithms, namely

K-Nearest Neighbor, (KNN) Gaussian Nave Bayes and Decision Tree Classifier for

forecasting and classification. (Goodwin P et al., 2002) review research on effective

approaches designed to allow judgment and statistical methods to be integrated when short-

term point forecasts are required. They pointed out that when forecasts are to be made, that

two approaches to the integration of judgment and statistical methods are usually identified.

Their paper identified that with regards to voluntary integration, the judgmental forecaster is

provided with details of the statistical forecast and then decides how to use this in

formulating their judgment. Thus the forecaster is free to ignore or accept the statistical

forecast when making the judgmental forecast. However, voluntary integration entails

applying judgmental adjustments to statistical forecasts. Nonetheless, it might also involve

17
the forecaster to modify “prior” judgmental forecasts should there be a newly arrived

statistical forecast. Whereas in mechanical integration, the integrated forecast is achieved by

applying a statistical method to the judgmental forecast.

(Gallagher C et al., 2015) research paper considers Bayesian classifiers to evaluate whether

they deliver higher performance for the task of sales forecasting for detecting deals that are at

risk of being lost or deals with the possibility of being won, which may have been classified

wrongly by Qualitative Sales Predictor (QSP). By combining both quantitative and

qualitative sales attributes to improve performance further, they used a more sophisticated

algorithm than those used in QSP. Their research claimed that Bayesian classifiers were the

most useful and suitable learning algorithms for the application and data.

More precisely, the TAN classifier had the maximum predictive accuracy in both testing and

validation. The final result produced 90.6% accuracy, implying that the TAN algorithm

drastically improved upon the QSP, which had an accuracy of 75.6% on the same test data.

Although data processing methods such as machine learning, having been developed in the

field of information science and are currently being employed in practical fields such as

medical data mining. These data mining techniques, alongside big data, are an area of great

interest. Researchers to date continue to discover knowledge and opinions that are ever

beneficial to businesses and organizations from the large amounts of customer information

data gathered by these businesses using these techniques. Among these, the deep learning

approach is a relatively new method of machine learning; though commonly and widely used

in image recognition, there have been little or no reports of its use in or instances of its

induction to marketing in retail businesses (Kaneko and Yada, 2016). Kaneko et al. (2016)

proposes a sales prediction model for retail stores as a way to introduce the deep learning

approach into the field of marketing so that store managers can use it for marketing strategies.

With demand forecasting being one of the main decision-making responsibilities of a

18
business, business intelligence (BI) becomes inevitable for analytics and strong decision

making for all sorts of businesses across all sectors. Thus, by analyzing past and current

market data, businesses and organizations can effectively predict the future demand of goods

and manufacture those goods that appear to have more demand in the future. In their work,

Khan et al. (2013) built a demand forecasting model structured on business intelligence and

machine learning.

Their model yielded a 92.38% accuracy when implemented in stores.

In 2015, (Rajab S et al., 2015) performed a comparison of prediction performance of two

commonly used AI methods, namely Adaptive Neuro-fuzzy inference system (ANFIS) and

Artificial neural network (ANN) for sales forecasting and stock market price prediction. The

results they obtained indicate that both ANFIS and ANN are indeed effective tools that can

be used for forecasting problems. Nevertheless, they pointed out that the learning period of

ANFIS is briefer, suggesting that ANFIS attains the target earlier than ANN. Additionally,

the accuracy they got for the ANFIS algorithm for the stock market dataset was greater than

ANN, which goes to show that ANFIS is a superior forecasting model. Whereas for sales

data, the ANN algorithm performed better. As noted by Islek et al. (2015) in their research

paper, demand forecasting is building forecasting models to estimate the quantity of a product

that customers will purchase in the future. Their work addresses the challenge of forecasting

several product demands of distribution warehouses. They presented a suitable methodology

for demand forecasting to overcome the limitations concerning product demand and

distribution in warehouses while providing a high estimation accuracy. Their proposed

methodology clusters similar warehouses according to their sales behavior and pattern by

employing bipartite graph clustering. They also introduced a hybrid forecasting phase that

combines the moving average model and Bayesian Network machine learning algorithms.

Another interesting research was carried out by Chen T et al. (2018), where they

19
implemented trend alignment with dual-attention multi-task recurrent neural networks for

sales prediction. They propose an innovative dual-attention based LSTM decoder to address

the issue by labeling the prominent factors in sales time series into internal features and

external features and then model them into two aspects with a multi task based LSTM

encoder. Their methodology allows the encoder to model various variables with different

semantic meanings expansively. With regards to multi-source learning, Tsai Ket al. (2020)

proposes a novel multi-source learning framework for sales prediction, allowing e-retailers to

preserve their products smartly and adjust the price appropriately to maximize profit.

XGBoost is used for binary classification in a supervised machine learning system capable of

matching internet devices to web cookies. XGBoost is used to improve the marketer’s ability

to identify individual users as they switch between devices and show relevant

content/recommendations to users wherever they go. Support Vector Machine (SVM) had

been applied to demand forecasting. Garcia et al. (2012), in their study, proposed an

intelligent model that relies on supporting vector machines to deal with issues relating to the

allocation and revelation of new models. Kandananond (2012) showed that SVM surpassed

Artifificial Neural Networks in estimating demand for consumer goods. Holmberg (2017)

observed that regular restaurant sales be influenced by the weather. They considered two

Machine Learning algorithms as XGBoost and neural network, and the results showed that

the XGBoost algorithm is more accurate than the other algorithm, and they also found that

they had improved their model performance by 2-4 percentage points by considering weather

factors. To improve accuracy, they had considered numerous variables such as date

characteristics, sales history, and weather factors.

Grigorios (2016) had used Machine Learning techniques to perform a survey on the

forecasting of food sales. They had addressed data analyst design decisions such as temporal

granularity, output variable, and input variables in this survey. In this paper, the authors

20
experimented by taking the point of sale (POS) as internal data and even external data by

considering different environments to enhance the efficiency of demand forecasting. They

considered different Machine Learning algorithms such as Boosted Decision Tree

Regression, Bayesian Linear Regression, and Decision Forest Regression for evaluation, In

particular, Support Vector Machine (SVM) had been applied to demand forecasting. Garcia

et. al., (2012), in their study, proposed an intelligent model that relies on supporting vector

machines to deal with issues relating to the allocation and revelation of new models.

Kandananond (2012) showed that SVM surpassed Artificial Neural Networks in estimating

demand for consumer goods.

XGboost is being used in research areas of growing importance such as miRNA-disease

association prediction, a problem that has been addressed using either Network Analysis or

Machine Learning techniques (Xing et. al., 2018). Also applied in old but continuing

problems such as bankruptcy prediction (Maciej et. al., 2018). In addition to other prediction

problems such as crude oil price (Mesut and Mustafa, 2017), grid-connected photovoltaic

plant production (Ferlito et. al., 2017), and as a part of a forecasting system for bird’s

migration (Benjamin & Kyle, 2018). Previously, most of the studies focused on considering

the metrics as mean absolute error, mean squared error, median absolute error, and k-fold

cross-validation is used for training and testing data. Metrics like max error, accuracy, and

mean absolute error are considered in this research. In this study stratified, the K-fold cross-

validation technique is used for training and testing to increase the eficiency of the results. In

this study, a suitable algorithm is chosen for sales forecasting.

Ali and Vehmas (2016) incorporate weather data into retail sales forecasting. They predict

daily sales of products based on time series forecast and a regression model accounting for

the weather effects. They find that forecasting accuracy for weather-sensitive product

categories can be improved by using this method. Fildes et. al., (2019) listed several machine

21
learning and nonlinear regression models used for sales prediction. They opined that most

published research has found improvements from using nonlinear methods rather than linear

regressions but remark that this may well arise from publication bias. The models used for

product sales prediction include back-propagation neural networks, fuzzy neural networks,

regression trees, grey relation analysis and multilayer functional link networks, two-level

switching models that select between a simple moving average and k-nearest neighbour or

decision trees, support vector machines (Pillo et. al., 2016), wavelets (Michis, 2015), fuzzy

set approach (Veiga et. al., 2016) and Bayesian P-splines (Lang e.t al., 2015). Also, extreme

learning machines are used in several studies.

Due to computational limitations, studies with nonlinear regression models often use only

small datasets (i.e., datasets with tens of products). The exception to this is Mukherjee et. al.,

(2018), where authors develop a demand density forecasting model with a multi-layered

hybrid deep neural network for inventory management in Flipkart.com. They model demand

density using a mixture of Gaussian distributions and report significant improvement in

prediction accuracy.

2.3. RESEARCH GAP

Potential research gaps arose from areas where existing literature or solutions were

insufficient or inadequate to address specific challenges or requirements. Here are some

potential research gaps:

• Integration of Unstructured Data:

Existing research often focuses on structured sales data, but incorporating unstructured data

sources such as customer reviews, social media sentiment, or competitor actions could

provide valuable insights for improving forecasting accuracy.

• Dynamic Model Adaptation:

22
Research on dynamic model adaptation techniques that can automatically adjust forecasting

models in response to changing market conditions or business dynamics is limited.

Developing adaptive models that can continuously learn and evolve over time could fill this

gap.

• Multi-Modal Forecasting:

Most existing forecasting models focus on univariate time series data, but in practice,

businesses may have access to multiple data modalities (e.g., sales data, inventory levels,

external factors). Research on multi-modal forecasting techniques that can effectively

integrate and leverage diverse data sources is needed.

• Quantifying Uncertainty:

While forecasting models provide point estimates of future demand, quantifying uncertainty

and providing probabilistic forecasts remains a challenge. Research on uncertainty estimation

techniques, such as Bayesian methods or ensemble approaches, could improve decision-

making under uncertainty.

• Interpretable Deep Learning Models:

Deep learning models, while effective for complex pattern recognition, often lack

interpretability, making it challenging to understand the underlying factors driving

predictions. Research on interpretable deep learning architectures or post-hoc explanation

techniques could enhance model transparency and trustworthiness.

• Supply Chain Dynamics:

Incorporating supply chain dynamics, such as lead times, production constraints, and

inventory policies, into forecasting models is essential for accurate demand prediction.

Research on supply chain-aware forecasting techniques that consider end-to-end supply chain

processes could fill this gap.

• Hierarchical Forecasting:

23
Businesses often need to forecast demand at multiple levels of aggregation, such as product

categories, regions, or customer segments. Research on hierarchical forecasting methods that

can effectively reconcile forecasts at different levels while capturing cross-level dependencies

is needed.

• Real-Time Forecasting:

Most forecasting models operate on historical data and require periodic updates, but real-time

forecasting is becoming increasingly important in dynamic business environments. Research

on real-time forecasting algorithms and systems that can provide timely and accurate

predictions is warranted.

• Ethical and Social Implications:

Limited research addresses the ethical and social implications of automated forecasting

systems, such as biases in data or decision-making, privacy concerns, and the impact on labor

markets. Exploring the ethical dimensions of forecasting and developing frameworks for

responsible forecasting practices could address this gap.

24