Incorporate machine learning in analyze amazon sales datasets to improve operational strategy

Ong, Qi Hao (2025) Incorporate machine learning in analyze amazon sales datasets to improve operational strategy. Final Year Project, UTAR.

PDF Download (3684Kb)

Abstract

This project is in the field of data analytics and machine learning, focusing on developing a short-term e-commerce sales forecasting system. The problem addressed is the difficulty for online retailers to predict sales trends due to seasonality, promotion-induced peaks, and irregular demand, and thereby leading to ineffective inventory planning and operations [2][3][4][13]. In order to overcome these challenges, three prediction models were used: Bidirectional Long Short-Term Memory (BiLSTM), Temporal Convolutional Network (TCN), and XGBoost. The study was carried out following the CRISP-DM methodology, beginning with data preprocessing and feature engineering from an openly accessible Amazon sales dataset (2017–2020) [28]. The data were condensed to daily sales figures, cleaned and converted, and subsequently modeled. BiLSTM modeled nonlinear temporal dependencies using a moving average, TCN learned long- and short-term patterns using causal convolutions, and XGBoost utilized lag features, rolling statistics, and calendar effects for interpretable tree-based forecasting. The models were evaluated using four metrics: R², RMSE, MAE, and MAPE. Results indicated that BiLSTM yielded the most balanced and accurate predictions (R² = 0.8481, MAPE = 12.91%), TCN had the highest explanatory power (R² = 0.9336, MAPE=24.35%) but overestimated peaks, whereas XGBoost performed poorly (R² = 0.6586, MAPE = 70.62%) despite being interpretable. To enhance practical adoption, an interactive Streamlit dashboard was developed, enabling users to upload sales data, select models, visualize forecasts, and receive AI-driven business insights. The novelty of this work lies in the combination of cutting-edge deep learning models and a decision-support dashboard that bridges the gap between predictive modeling and actionable strategy. In short, the system produces reliable short-term forecasts and interpretable recommendations, thereby forming an effective tool for operational and strategic planning for e-commerce.

Actions (login required)