Time-series foundation models (TSFMs) can impute missing values and forecast unseen series without training, but the effect of zero-shot imputation on a subsequent zero-shot forecast has not been measured. This paper evaluates a fully zero-shot impute-then-forecast pipeline. Three TSFM imputers (MOMENT, TSPulse and TS-ICL) and linear interpolation fill the contexts of five public datasets from which 4 to 32 percent of all values have been removed, either uniformly at random or concentrated in the most recent part of each series (tail-concentrated), and two TSFM forecasters (Chronos-2 and TimesFM 3.0) forecast from each imputed context. Imputation is causal, and each forecast is compared with a forecast from the clean context. Under random missingness, forecasting accuracy on four datasets stays close to the clean-context reference for every imputer except MOMENT, whereas under tail-concentrated missingness, which leaves about half of the recent context missing, every imputer degrades it on average. TS-ICL is the most accurate and least harmful imputer under both patterns, whereas linear interpolation is competitive only under random missingness.

time-series foundation models,missing data imputation,zero-shot forecasting