Improving Automotive Production Forecasts via Explainable AI-Driven Feature Pruning

Abstract This research uses explainable AI (XAI) to improve machine learning (ML) models in forecasting production by identifying and pruning low-value sensor-derived features in production forecasting tasks. The methodology entails the initial training of ML models, followed by a fine-tuning phase in which irrelevant features identified through explainability methods are eliminated. It builds on the throughput prediction architecture of BMW by performing XAI-driven refinement on top of an existing Extra Trees ensemble framework. Using Shapley values and permutation importance, redundant, unstable, or low-signal inputs are removed. This reduces overfitting, improves model interpretability, and compresses inference time without sacrificing accuracy. The result is both performance improvements and simplified model logic, enabling potential reductions in manufacturing costs and improved model transparency. Attribution monitoring also enables early detection of deteriorating predictive logic—supporting more transparent, auditable, and adaptable decision-making in automotive manufacturing environments.