RateIN: Time Series Foundation Models Have a Sampling Problem

A zero-shot forecaster receives a series at whatever sampling rate it was recorded at, and nothing guarantees thatthis rate suits the model. We present RateIN, an inference-time wrapper that resamples the context by a factor k,lets a frozen model forecast on the new grid, and brings the quantile forecast back to the native grid. The factoris chosen once per configuration (a dataset at one sampling rate and horizon) by a backtest: candidate factors aretried on the recent past of its series and the best ones are mixed. On GIFT-Eval, RateIN lowers the CRPS ofall six models we tested, from 1.4M to 256M parameters, by 0.5 to 11 % against the same model with a sign-flipaugmentation (the average of its forecasts for x and −x). The gain is clear for both Chronos-Bolt models andTimeSSM-2.5M; for Chronos-2, t0-beta and TTM-r3 its 95 % bootstrap interval reaches zero. Where it acts, it actsthe same way for every model: on series recorded at 96 to 360 samples per cycle, the best single factor brings themdown to about 30 samples per cycle, and the backtest’s mixture to about 60. Controls show that the gain is not asmoothing effect: for four of the six models most of it comes from the coarser rate itself, and for TimeSSM-2.5Mfrom the longer history that decimation brings into the window. A learned selector that sets k from the contextwindow makes TimeSSM-2.5M worse: the best k cannot be read off one window. The backtest costs about 11times the forward passes of the augmented model. A fixed rule that moves every series to 24 samples per cyclefrom its declared season costs nothing extra, matches the backtest on both Chronos-Bolt models and keeps 73 %of its gain on t0-beta. Measuring once how many samples per cycle each model prefers, then moving series there,would make RateIN cheap. Code: https://github.com/IUseAMouse/TimeJEPA (tag ratein-paper-v2) and https://github.com/IUseAMouse/TimeSSMWeights: https://huggingface.co/IUseAMouse/timessm-2.5m