Data Evaluation Framework for Time-Series Foundation Models

Foundation models are trained on large, heterogeneous collections of time series, but not every sample earns its place. This work selects an optimal pretraining subset by characterizing data in a low-dimensional space built from temporal, statistical, and spectral attributes together with foundation-model embeddings, relating neighborhood structure to downstream forecasting quality.

Overview

Problem — pretraining data mixtures are usually chosen by intuition, with little principled signal on what each sample contributes.
Approach — a KNN-based selection algorithm identifies influential and diverse samples; robust attribute estimators and embedding extractors handle ill-defined statistics and outliers.
Result — comparable forecasting performance with a substantially smaller pretraining corpus, tested across foundation-model architectures. Published at ICASSP 2026.