Publication Details
Overview
 
 
Nadir Ehmimed, Mohamed Yassin Chkouri, Abdellah Touhafi
 

Chapter in Book/ Report/ Conference proceeding

Abstract 

The practical robustness of time-series forecasting models on imperfect sensor data remains a critical issue. This paper challenges common heuristics by rigorously comparing two paradigms for achieving robustness: the complex architectural path (Transformer) and the simpler feature engineering path (XGBoost). Using matched hyperparameter search budgets to ensure a fair comparison, our results illuminate a fundamental trade-off between performance consistency and computational efficiency. The well-tuned Modern Transformer achieves the most consistent performance, resulting in the best average MAE of 0.677. In contrast, the feature-based pipelines perform exceptionally well on most data, but are susceptible to a distinct high-error failure mode, resulting in a mean MAE of approximately 0.78. Crucially, we find that this small, statistically insignificant difference in average performance comes at a steep price: the feature-based path is orders of magnitude more computationally efficient. We conclude that the choice of paradigm is not about universal superiority, but a strategic decision. While well-tuned architectures offer the most consistent path to state-of-the-art accuracy, the feature engineering paradigm provides a vastly more efficient baseline. This work provides an empirically-grounded guide to navigating the critical trade-offs between performance and efficiency for real-world forecasting systems.

Reference