Foundation ModelGoogle Research

TimesFM

Decoder-only foundation model pre-trained on ~100B real-world time points. Patched-decoder attention generalises across history lengths, horizons, and frequencies; zero-shot performance closes the gap to fully supervised baselines at a fraction of the parameter count.

TimesFM is Google Research's decoder-only foundation model for zero-shot forecasting. The architecture is a patched-decoder transformer: history is split into patches, each patch is embedded and attended over autoregressively, and the head predicts the next patch.

Pretraining mixes ~100B real-world time points from public sources with synthetic series. The single model generalises across granularities from minutes to years and across history/horizon combinations, so the same checkpoint can serve every challenge on this platform without per-task adaptation.

Two versions are evaluated here. The 2.0 release ships a larger 500M-parameter model; the 2.5 generation trades size for richer pretraining data, landing at 200M parameters with stronger reported zero-shot accuracy on GIFT-Eval.

Versions on TS-Arena

Each version below corresponds to one registered model id in the leaderboard. Click through to its detail page for per-model rankings, forecasts, and history.

  • TimesFM 2.0 (500M)
    timesfm-2.0-500m
    500M params

    Largest TimesFM checkpoint. Trained on the TimesFM 1.0 dataset + a LOTSA subset.

  • TimesFM 2.5 (200M)
    timesfm-2.5-200m
    200M params

    Latest generation. Smaller but trained on the expanded TimesFM 2.0 dataset.