A Comparative Study of AutoRegressive Integrated Moving Average, Gradient Boosting, and Hybrid models for Inventory Levels Forecasting in Retail Supply Chains

Abstract

Accurate​‍​‌‍​‍‌ inventory forecasting is essential in retail supply chains that deal with highly volatile and non-linear demand situations. This research compares the performance of three forecasting models ARIMA, Gradient Boosting Regressor (GBR), and a hybrid ARIMA+GBR model using a dataset of 72,400 enriched daily observations. The models are tested with an 80/20 time-series split and several metrics (MSE, MAE, RMSE, R², WAPE, sMAPE).  

The results show that ARIMA is the least performant method, while the Hybrid model helps to stabilize the results but does not outperform GBR. The GBR model achieves the best results as it has the lowest errors and the highest R² value. Robustness tests also show that the Hybrid model can handle noise and is still stable, however, it does not outperform the GBR.

The study confirms the effectiveness of machine learning approaches, particularly GBR, for short-term inventory forecasting and suggests future research directions, such as ARIMAX and deep learning models.