XGBoost and Time series Model Fusion for Sales Trend Forecasting on E-commerce Platforms

Abstract

Targeting the e-commerce setting characterized by pronounced demand volatility and complex exogenous drivers, this study addresses the accuracy and robustness of multi-horizon trend forecasting. We note that classical statistical models tend to destabilize when absorbing high-dimensional covariates and under promotion/holiday shocks, while pure machine-learning and deep models often lack seasonality priors and interpretability. To this end, we build a fusion framework of "time-series decomposition - residual learning - lightweight stacking": SARIMA/Prophet provides a trend-seasonality baseline; XGBoost learns nonlinear residual responses to exogenous variables; and ridge regression adaptively fuses the two branches across horizons. Using two years of daily, multi-category data, we conduct rolling-origin evaluation and a final holdout test. Results show that, relative to a strong non-fusion baseline (XGBoost), our approach reduces MAPE and RMSE by approximately 16.5% and 9.7%, respectively, and markedly compresses peak errors within promotion/holiday windows, while maintaining CPU-level low-latency inference and stable production monitoring. The contributions lie in a unified paradigm that combines structural priors with nonlinear expressiveness; an event-weighted training objective and robust feature encoding tailored to promotion shocks; and interpretability analyses that reveal the pivotal roles of promotion intensity, relative price, and traffic in improving forecasts - thereby providing quantifiable guidance for pricing, inventory planning, and advertising allocation in e-commerce operations.