GITCO boosts TimesFM 2.5 accuracy by 1.95% with context optimization
New gated framework fixes context poisoning without retraining models.
A team of researchers from multiple institutions has introduced GITCO, a novel inference-time method to boost the accuracy of patch-based Time Series Foundation Models (TSFMs). These models suffer from "context poisoning," where anomalous input patches grab disproportionate attention and silently reduce forecast quality. Instead of retraining or fine-tuning model weights, GITCO optimizes the input context at inference time using three lightweight components: a Gate that scores each patch's relevance, a Router that selects which patches to prune, and a Critic that validates the final context.
Evaluated on Google's TimesFM 2.5 across 53 diverse GIFT-Eval datasets using K-fold cross-validation, GITCO achieved an average 1.95% reduction in MASE (a standard forecast error metric) while capturing 89.9% of the theoretical improvement upper bound. The paper also introduces "context sensitivity profiles" — a new way to characterize how a TSFM's architecture and the statistical properties of data jointly determine the expected accuracy gain from context intervention. This opens the door for adaptive, model-specific optimization strategies without costly retraining.
- GITCO uses Gate, Router, and Critic modules to prune harmful patches in TSFM inputs without any parameter updates.
- Outperforms baseline TimesFM 2.5 by 1.95% average MASE reduction across 53 GIFT-Eval datasets.
- Captures 89.9% of the theoretical upper bound for inference-time improvement.
Why It Matters
GITCO offers a lightweight, training-free way to improve time series forecasts, critical for finance, energy, and IoT applications.