Abstract

This study evaluates short-term forecasting of chlorophyll-a (Chl-a) dynamics at two monitoring stations in Lake Villarrica, northern Patagonia, Chile, using reproducible station-level Chl-a time series and foundation timeseries models. The forecasting experiments used univariate Chl-a context windows as inputs to TimesFM and Chronos Large and evaluated 1-30 day horizons under a rolling-origin design. Short-horizon skill was highest at the 1-day horizon: at La Poza, D1 MAE was 0.099 & micro;g/L for TimesFM and 0.106 & micro;g/L for Chronos Large; at Pucon, D1 MAE was 0.063 & micro;g/L for TimesFM and 0.071 & micro;g/L for Chronos Large. Forecast error increased with lead time, with D28 MAE ranging from 0.606 to 0.757 & micro;g/L across stations and models. Lag diagnostics showed that most D7 forecast cases had maximum correlation at lag 0. However, Pucon Chronos Large showed a non-zero best lag, indicating that peak timing remains uncertain for warning applications. We also evaluated q10/q50/q90 predictive intervals and observed/gap-stratified performance. The results support the use of foundation time-series models for short-term forecasting of reconstructed Chl-a products, while highlighting limitations due to imputation, missing exogenous drivers, and the absence of validated satellite-feature ingestion in the current forecast architecture.