an integrated machine learning framework for retail sales forecasting

Sales prediction needs to be accurate for effective inventory management, supply chain planning, and data-driven decision-making in the retail sector. However, traditional statistical forecasting methods often struggle to model the complex and dynamic patterns that exist in modern retail sales, while the predictive accuracy of machine learning models varies across datasets and market demand settings. This study developed and evaluated an integrated machine learning framework for retail sales forecasting using historical sales data obtained from Supermarket S. The framework integrated data cleaning, exploratory data analysis, outlier investigation, feature engineering, standardization using StandardScaler, model development, and comparative performance evaluation. Machine learning and statistical forecasting techniques, such as ARIMA, Exponential Smoothing, Random Forest, Gradient Boosting, XGBoost, and SVR, were investigated. Additionally, ensemble strategies were explored to determine whether combining the best machine learning models could further boost forecasting performance. A validation-weighted RF-GB-XGBoost ensemble was developed by assigning model weights derived from validation forecasting residuals. MAE, MAPE, WMAPE, and R² were used to test model robustness. The findings revealed that MAPE was disparately affected by low sales observations within the highly dispersed sales distribution; therefore, WMAPE was included to give a more representative volume-weighted assessment of forecasting errors. XGBoost yielded the best overall forecasting performance, with an MAE of 35,196.02, WMAPE of 13.96%, and R² of 0.8362. The validation-weighted ensemble attained an MAE of 45,835.67, WMAPE of 18.18%, and R² of 0.7407 and did not outperform XGBoost alone. The results illustrate that integrating multiple forecasting models does not automatically improve predictive accuracy, particularly when weaker base learners reduce the contribution of the strongest individual model. To test the robustness and adaptability of the proposed framework, an additional validation experiment was conducted using the publicly available Walmart retail sales dataset. The study concludes that systematic model comparison and appropriate evaluation metrics are important for selecting reliable forecasting techniques.