![[2602.22520] TEFL: Prediction-Residual-Guided Rolling Forecasting for Multi-Horizon Time Series](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.22520] TEFL: Prediction-Residual-Guided Rolling Forecasting for Multi-Horizon Time Series
Summary
The paper presents TEFL, a novel framework for multi-horizon time series forecasting that utilizes prediction residuals to enhance accuracy and robustness in forecasting models.
Why It Matters
TEFL addresses critical challenges in time series forecasting by incorporating historical prediction residuals, leading to improved accuracy and resilience against abrupt changes. This advancement is significant for industries reliant on accurate forecasting, such as energy and transportation.
Key Takeaways
- TEFL integrates past prediction residuals into forecasting models.
- The framework improves accuracy by reducing MAE by 5-10% on average.
- TEFL demonstrates robustness against abrupt changes and distribution shifts.
- A two-stage training procedure optimizes both the forecaster and error module.
- Extensive testing across multiple datasets validates TEFL's effectiveness.
Computer Science > Machine Learning arXiv:2602.22520 (cs) [Submitted on 26 Feb 2026] Title:TEFL: Prediction-Residual-Guided Rolling Forecasting for Multi-Horizon Time Series Authors:Xiannan Huang, Shen Fang, Shuhan Qiu, Chengcheng Yu, Jiayuan Du, Chao Yang View a PDF of the paper titled TEFL: Prediction-Residual-Guided Rolling Forecasting for Multi-Horizon Time Series, by Xiannan Huang and 5 other authors View PDF HTML (experimental) Abstract:Time series forecasting plays a critical role in domains such as transportation, energy, and meteorology. Despite their success, modern deep forecasting models are typically trained to minimize point-wise prediction loss without leveraging the rich information contained in past prediction residuals from rolling forecasts - residuals that reflect persistent biases, unmodeled patterns, or evolving dynamics. We propose TEFL (Temporal Error Feedback Learning), a unified learning framework that explicitly incorporates these historical residuals into the forecasting pipeline during both training and evaluation. To make this practical in deep multi-step settings, we address three key challenges: (1) selecting observable multi-step residuals under the partial observability of rolling forecasts, (2) integrating them through a lightweight low-rank adapter to preserve efficiency and prevent overfitting, and (3) designing a two-stage training procedure that jointly optimizes the base forecaster and error module. Extensive experiments across 10 re...
