Sales Forecasting for a Global Superstore Using Time Series and Machine Learning Models

Abstract

This study is based on the Global Superstore dataset and conducts a comparative analysis of the performance of traditional time series models and machine learning models in retail sales fore casting. The study adopts the CRISP-DM methodology and evaluates the predictive performance of different models based on two time granularities: daily sales and weekly sales. The models used in the research include ARIMA, Holt-Winters exponential smoothing model, lin ear regression, random forest, XGBoost, and artificial neural network (ANN). For the machine learning models, the study further constructed feature engineering methods such as temporal fea tures, lag features, and rolling features. The performance of the models was evaluated using mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), explained variance score, and maximum error. The study results show that in both daily sales forecasting and weekly sales forecasting, the linear regression model using the complete feature set (Full Features) performed the best overall, while more complex models did not significantly outperform the simple models. Compared to daily sales forecasting, weekly sales forecasting is more stable overall due to the reduction of noise caused by time aggregation. However, all models still have obvious limitations in predicting abnormal sales peaks and extreme fluctuations. The research results indicate that in retail sales scenarios with limited data scale and significant fluctuations, models with relatively simple structures and strong interpretability can still provide certain practical predictive value.