t0: A Time-Series Foundation Model for Forecasting with Context
The Forecasting Company has released t0, a family of open-weights foundation models for forecasting with multivariate context, together with a technical report on arXiv. Arnau Quera-Bofarull and I are co-authors for Macrocosm. We contributed Section 5.6.2, an independent evaluation of the models on electricity prices in Texas.
Read it on arXiv · Weights: t0-alpha · Weights: t0-beta · Code
The first two models are t0-alpha (102M parameters) and t0-beta (256M). They forecast from the history of the target series, past covariates and known-future covariates, without task-specific retraining, and return probabilistic forecasts as quantiles. The transformer layers alternate attention along time and across variates. Pretraining combines curated public data with synthetic generators built to contain covariate-to-target dependencies.
Our section covers a back test on hourly real-time electricity prices in four ERCOT load zones, from January 2024 to May 2026. Each model forecasts the next operating day from the information available at 09:00 Central Time the day before, including 21 covariates for load, renewable generation, outages and weather. Both models cut the mean absolute error of a lagged-price baseline by more than 38%. TimesFM-3 has the lowest error and CRPS among the pretrained models compared, and the 80% prediction intervals of the t0 models contain 74 to 75% of observed prices: they cover some peaks and miss the extreme spikes. The section also converts the forecasts into trading rules, selected retrospectively for each model, and reports the resulting profit and Sharpe ratios.
The architecture, training and the remaining evaluations are in the report.