?
Automated Feature Engineering Based on Explainable Artificial Intelligence for Time Series Forecasting
This work presents a practical, explainability-guided pipeline for time-series forecasting that integrates automated lag engineering, XAI-based feature selection, and a lightweight, post-hoc calibration of a tree-ensemble forecaster. Rather than proposing a new forecasting paradigm, we show that FI-SHAP explanations stabilized by global feature-usage can flag redundant lag features for removal, and an exponential-smoothing–anchored calibration of LightGBM (ES–LightGBM) can mitigate mean-level bias and trend-extrapolation limits in some regimes. Evaluated on four public datasets under multi-step settings, the resulting pipeline is competitive with representative deep-learning baselines under the evaluated conditions: in most configurations it attains the lowest or tied-for-lowest error and provides an additional 1%–8% MSE reduction over the strongest baseline considered. The approach offers transparent feature rationales and minimal compute overhead, highlighting how XAI can make a standard tree ensemble both interpretable and practically strong for multi-horizon forecasting.