Bayesian non-parametric method for decision support: Forecasting online product sales

Abstract

Forecasting online product sales is essential for retailers and e-commerce platforms, but it can be challenging owing to the complex dynamics and mixed trends in sales data. Popular end-to-end approaches tend to capture spurious correlations in historical data, whereas two-stage approaches that decompose time series and make separate predictions often result in error accumulation. To address these issues, we propose PoissonGP, a novel Bayesian model that employs a non-homogeneous Poisson process with a Gaussian process prior for sales prediction. PoissonGP can capture complex patterns in data with multiple trends and manage distribution shifts caused by changes in long-run sales. Additionally, it incorporates forecast uncertainty and provides interpretability for building efficient and visual decision support systems. Experimental results on several synthetic and empirical datasets suggest that PoissonGP outperforms existing approaches, making it a promising tool for e-commerce platforms.