A Comparative Analysis of Statistical and Machine Learning Methods for Retail Demand Forecasting to Support Operational Planning

Abstract

Retail demand forecasting is essential for businesses in their operational planning and supply chain management. Since the rise in the level of competitiveness in the markets and the adoption of data-driven decisions is high, there is a growing need for forecasting in businesses. For retailers, forecasting is essential because it aids in planning the workforce, distribution, reorder levels, and inventories. In contrast to the conventional statistical approaches that have dominated forecasting, the machine learning algorithms have introduced alternative models for complex demands. There is, however, little certainty about the relative effectiveness and viability of these forecasting approaches in the retail environment.

This study aims to compare statistical and machine learning techniques to retail demand forecasting and the relevance of the approaches in retail operational planning. The research will compare the accuracy of the forecasted values, the behaviour of the models, and how helpful the approach is to the decision-making process in the retail sector.

This research work has been conducted based on quantitative and empirical techniques. The empirical analysis has been conducted based on publicly available sales transaction data obtained from the UCI Machine Learning Repository. The dataset has been pre-processed through data cleaning, data aggregation, outlier handling, and feature engineering processes. In addition to that, the fixed-origin-hold-out validation technique has been used for model comparison for the Naive, ETS, SARIMA, XGBoost, Random Forest and Neural Network models. Forecasting accuracy has been calculated using forecasting accuracy metrics MAE, RMSE, and WMAPE.

The findings from the analysis prove that machine learning approaches outperform conventional statistical methods in managing the data with high volatility in the retail industry. From the results obtained, Random Forest emerged to be the most reliable method in generating unbiased predictions as it performs well in balancing bias and variance. The study shows that precise forecast results could improve operational planning through reduced risks while decision-making on inventory management. The research is an addition to the existing forecasting and retail analytics body of knowledge through employing a holistic analysis approach using public retail datasets to benchmark various forecast methods. Future studies are recommended to incorporate external exogenous factors in the prediction model along with a hybrid architecture system.