Get Started
Home
Topics
Search
Library
Research questionHow can historical base-model residuals robustly correct multivariate forecasts without overcorrection?Historical target values may be unreliable correction signals when contexts differ in level, scale, or local dynamics. Base-model residuals can capture recurring forecast errors, but forecasting systems do not typically retain and retrieve those individual errors for later correction.
AI
AI Memory
Machine Learning
Retrieval-Augmented Generation
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.RATL: Learning from Retrieved Residuals for Robust Multivariate Time-Series ForecastingThe source studies a plug-in residual-memory approach for frozen base forecasters. It builds a train-only memory of historical residual trajectories, retrieves residuals from similar contexts subject to causal availability, and combines them across forecast blocks and variables; correction strength is selected using validation. Evidence comes from real-world benchmarks using iTransformer as the primary backbone, comparisons with other forecasting baselines, ablations, and transfer tests across backbones.research paper · Sep 3, 2026
Related questions
How can reinforcement-learning corrections adapt a global weather model without destabilizing its dynamics?How can diffusion models forecast multivariate time series reliably when variables differ in noise and observations are corrupted?How can online time-series forecasters adapt to evolving, unseen patterns without costly continual parameter updates?How can multivariate time-series imputers recover structured gaps under high missingness?