Beyond Valuation Scores: Data Selection for Time-Series Foundation Models

Time-series foundation models (TSFMs) can be adapted to historical data, but the value of that data and whether it transfers across models and future conditions remains poorly understood. We study data valuation as a mechanism for selecting adaptation data, separating the utility of adaptation itself from the utility of valuation-guided selection. Across three valuation methods, two pretrained TSFMs, and three forecasting datasets, we find that these two notions often diverge: adaptation may help or hurt independently of whether a selected subset outperforms random selection. Selection utility is also strongly dependent on the model, budget, dataset, and deployment period. In particular, effects that are substantial under TTM largely attenuate under Chronos-2. These results show that TSFM data valuation should be evaluated through end-to-end adaptation utility and portability, rather than valuation scores alone.