GRAM: Correcting Frozen Time-Series Foundation Models via Graph-Retrieved Amplitude Memory

2026-10-04

AI Summary

This study addresses the challenge of extracting correction signals from frozen time-series foundation models, where random noise interference and incomparable error scales hinder effective post-hoc refinement. To this end, we propose GRAM, a framework that introduces an amplitude-normalized error prototyping mechanism. Specifically, an amplitude memory module uniformly rescales and aggregates prediction errors, while a prototype graph neural network captures the relational structure among these errors. This design effectively suppresses stochastic fluctuations and precisely disentangles systematic biases to generate hour-level correction signals. As a zero-shot retrieval-augmented learning paradigm, GRAM achieves consistent and significant performance improvements across multiple datasets and diverse foundation models without requiring task-specific fine-tuning.

Abstract

Time-series foundation models (TSFMs) enable zero-shot forecasting through large-scale cross-domain pretraining, while retrieval augmentation further improves their performance by leveraging historical information. However, existing methods typically correct TSFM forecasts using the ground-truth futures of similar historical windows, which contain both predictive components already captured by the foundation model and sample-specific random fluctuation that is difficult to transfer. In contrast, recurring systematic model bias within prediction errors more directly characterizes the failure modes of a frozen TSFM and therefore provides more valuable correction signals. Effectively exploiting such model bias, however, poses two challenges: prediction errors at different numerical levels are difficult to compare due to scale differences, and the recurring bias must be extracted from prediction errors contaminated by random fluctuation. To address these challenges, we propose GRAM, a general retrieval-augmented framework for frozen TSFMs. GRAM first introduces an Amplitude Memory Module (AMM) that scales prediction errors by amplitude and aggregates them into retrievable prototypes. It then employs a Prototype Graph Module (PGM) to model relations among prototypes to aggregate consistent bias information while suppressing random fluctuation. During online forecasting, GRAM retrieves and expands prototypes relevant to the current query and generates per-horizon corrections to refine the original TSFM forecast. Experiments across multiple datasets and foundation models demonstrate consistent forecasting improvements.

Problem

• Time-Series Foundation Models
• Forecasting Correction
• Systematic Model Bias
• Retrieval Augmentation
• Prediction Errors

Innovation

• Time-Series Foundation Models
• Retrieval Augmentation
• Amplitude Memory Module
• Prototype Graph Module
• Bias Correction