Accuracy comparison between Sparse Autoregressive and XGBoost models for high-dimensional product sales forecasting

Abstract

[en] Predicting future sales is key for any business budgeting and resource allocation. One major concern when trying to build accurate forecasts are the cross-category relationships between some products and the effect that might have on each other’s sales. Given today’s data abundance, this issue is even more worrying: traditional statistic models can’t handle high-dimensional datasets with ten or more products. With the use of popular machine learning and data science tools, we developed a framework that enables the building, training and evaluation of two models and its comparison through a detailed set of forecast metrics 1 . The first model is a modified Vector Autoregressive model (VAR) which takes into account product relationships. The second one is an XGBoost model, which is not specialized into cross-category associations but it’s known for its versatility and performance when working with tabular data. After performing a one-month ahead sales forecasting on a huge dataset of multiple product sets, we find that inter-product connections play a huge role in prediction accuracy since the VAR model performed considerably much better than the XGBoost.