Time Series Regression for Multi-Store Sales Forecasting Using Machine Learning

Abstract

For retail companies, having accurate sales forecasting is essential for inventory management, financial allocation and operational decisions in geographically diverse stores. But traditional forecasting approaches may not be able to capture complex temporal relationships, the storelevel variations and the non-linear changes in sales demand. We develop a centralised time series regression system based on a multi-store retail data set, which includes historical sales data for the observation period, store specific features and temporal features. The data are preprocessed using data cleaning, missing-value treatment, chronological sorting, feature scaling and converting historical observations into lag based and calendar based features. Linear Regression, RF Regression, Gradient Boosting Regression, LSTM and GRU networks are constructed and compared to check the performance of standard ML and DL approaches. The performance of the models is tested by the MAE,RMSE, MAPE and coefficient of determination (R2). The experimental results demonstrate that DL models outperform in forecasting; the most optimal model has a higher accuracy in capturing the long-term temporal correlation and non-linear sales trends across multiple stores. The proposed centralised modelling approach is effective and scalable to enhance the accuracy of multi-store retail sales forecasting.