ECOMMERCE SALES FORECASTING USING MACHINE LEARNING ALGORITHMS IN

BANGLADESH

1.1 INTRODUCTION
In today's digital landscape, accurate sales forecasting is essential for e-commerce businesses in
Bangladesh to optimize operations and meet customer demand effectively. Leveraging past sales data,
this predictive technique empowers organizations to anticipate future trends, manage inventory
efficiently, and make informed strategic decisions.

Traditionally, sales prediction relied on a single model approach. However, with the advent of advanced
machine learning algorithms like XGBoost and Random Forest, a new era of forecasting accuracy has
dawned. These algorithms offer enhanced capabilities in analyzing large datasets, identifying patterns,
and predicting sales trends with precision.

This research focuses on employing both XGBoost and Random Forest algorithms to enhance sales
forecasting in the e-commerce sector of Bangladesh. Through rigorous data preprocessing and analysis,
the aim is to identify significant factors influencing sales performance. By comparing the performance of
these two algorithms, the study seeks to determine the most effective approach for predicting sales
accurately. Additionally, feature importance and permutation importance analyses will be conducted to
gain insights into the key variables driving sales predictions, further enhancing the accuracy and reliability
of the forecasting models. Moreover, by generating insights from these analyses, actionable
recommendations for e-commerce businesses in Bangladesh to optimize their sales strategies and
improve overall performance can be provided.

1.2 RESEARCH OBJECTIVE

The objective of this research report is to investigate and apply machine learning (ML) algorithms,
specifically Random Forest and XGBoost, to enhance e-commerce sales forecasting in the context of
Bangladesh. The primary goals are:

• Utilize advanced ML techniques to predict e-commerce sales accurately.

• Identify critical factors influencing e-commerce sales trends and patterns.
• Compare and analyze the effectiveness of Random Forest and XGBoost algorithms in sales
forecasting.
• Investigate feature importance to understand key drivers of e-commerce sales in Bangladesh.
• Provide actionable insights and recommendations for businesses based on the findings.

By achieving these objectives, this research aims to contribute to the optimization of sales forecasting
strategies and decision-making processes within the e-commerce sector of Bangladesh, ultimately
enhancing operational efficiency and market competitiveness.
 1.3 LITERATURE REVIEW

Similar Temporal RangeofData

Predictor Outcome Source Methodology Country Author Best Fit
Research
Paper Name
Sales Forecasting Price, Rating, Sales Google Scholar, Microsoft Machine Brazil Revista ANN
Using Number of Forecasting Academic,and Scopus Learning Gesec
MachineLearning Reviews databases in the period (ANN)Or
Algorithms from 2015 to Algorithmor Big
08/2021 Data
Analytics

Predicting Product Sales Sales Retail Store Chain Benchmark Model, Tuukka XGboost
SalesIn Retail Store Transactions Prediction Operating (Monthly Lasso Regression, Finland Sarvi
Chain and Related Sales) And XGboost
Dimension
Tables (E.g.,
Products And
Stores)

Machine Items Sales Chunks Ranging Recurrent Neural Unknown Austin Recurrent Neural
Learning Sold, Forecasting From8 September2016 Network (RNN) Schmidt Network (RNN)
Based Gross To 31 December 2016, Temporal Fusion
Restauran Sales, 1January 2017 To 31 Transformer (TFT)
tSales Payment December 2017, And1
Forecasti Information March 2018 To 5
ng edrfcDecember 2019

MRP, Future Sales Unknown XGBoost India VarshiniS XGBoost

An Analysis of Output Regressin,
Machine Learning Type, Artificial
Algorithms to Predict Item NeuralNetwork
Sales Type, (ANN),Support
Outlet Vector
Locatio Regression (SVR)
n
Sales Total Amountof Predicts Daily The Annapolis Cider Machine Learning, Canada Yuanyuan Persistence Model
ForecastingUsing Sales For Each and Hourly Company has provided all Artificial Neural Shi Avg6Week Model
Long Short Term Year SalesWithin sales transaction data Networks, Time
Memory Networks 10% Error stored in their POS system Series, LSTM and
for four years(April 2016 to Prior Methods
March Persistence Models,
2020). Avg6Week Model

Data-driven Total Order E-commerce Online Sales DataFrom Machine Learning Sweden Liubov LSTM
Models For E- Amount, Items Company’s TheClient. Algorithm LSTM Guseva
commerce Category Future Sales January2019 to
Sales September 2022
Predictions
Sales forecasting is crucial for businesses, and with machine learning (ML) techniques, it's become more
effective. "Sales forecasting using machine learning algorithms" highlights ML's efficacy in improving
accuracy by handling complex variables and identifying hidden demand patterns. Traditional methods
may suffice in stable markets. "Predicting product sales in retail store chain" shows ML algorithms like
XGBoost and Lasso regression add value, particularly for existing and new products. Further research is
needed for seasonal patterns and prediction accuracy. "Machine Learning Based Restaurant Sales
Forecasting" presents a scalable sales prediction model using recurrent neural network (RNN) models,
emphasizing the importance of weather data. "An Analysis of Machine Learning Algorithms to Predict"
finds Random Forest Regressor as the top-performing algorithm for retail sales prediction. "Analysis of
Machine Learning Model for Predicting Sales Forecasting" highlights the superiority of the Random Forest
algorithm in predicting revenue in the retail sector. "Data-driven models for e-commerce sales
predictions" compares LSTM, Prophet, and a custom ASP model, with LSTM emerging as the most
effective for workload management and logistics planning.

Research has shown ML algorithms like XGBoost and Random Forest (RF) excel in handling complex data
and predicting sales patterns. "Sales forecasting using machine learning algorithms" highlights the
effectiveness of ML algorithms in handling complex variables and identifying hidden demand patterns. In
e-commerce, these algorithms outperform others, aiding workload management and logistics planning.
Adapting these methods to the Bangladeshi e-commerce scene, considering local variables like consumer
behavior, market trends, and socio-economic factors, can improve forecasting accuracy and provide
valuable insights for businesses in the region.