Human-in-the-Loop Adaptive Optimization for Improved Time Series Forecasting

ICLR 2026 Conference Submission19278 Authors

Keywords: time series, forecasting, human-in-the-loop, post-training

TL;DR: This work introduces a human-in-the-loop optimization framework using adaptive techniques to leverage human expertise interactively, significantly improving forecasting accuracy

Abstract: Time-series forecasting models often produce systematic and predictable errors, even in critical domains such as energy, finance, and healthcare. We introduce a novel post-training adaptive optimization framework that improves forecast accuracy without retraining or architectural changes. Our approach adds a lightweight model-agnostic correction layer that automatically finds expressive output transformations optimized by reinforcement learning, contextual bandits, or genetic algorithms. Theoretically, we prove the benefit of an affine correction and quantify the expected performance gain together with its computational cost. The framework also supports an optional human-in-the-loop component: domain experts can guide corrections using natural language, which is parsed into actions by a language model. Across multiple benchmarks (e.g., electricity, weather, traffic), we observe consistent accuracy gains with minimal computational overhead. Our interactive demo (link) showcases the usability of the framework in real time. By combining automated post-hoc refinement with domain-expert corrections to the base forecasting model, our approach offers a lightweight yet powerful direction for practical forecasting systems.

Primary Area: learning on time series and dynamical systems

Submission Number: 19278