TimesFM-3

TimesFM-3 is a state-of-the-art, 330-million parameter foundation model designed for zero-shot multivariate time series forecasting. It leverages a decoder-only transformer architecture to jointly predict multiple coevolving time series while incorporating past and future covariates to improve accuracy.

TimesFM-3 is a state-of-the-art, 330-million parameter foundation model designed for zero-shot multivariate time series forecasting. It leverages a decoder-only transformer architecture to jointly predict multiple coevolving time series while incorporating past and future covariates to improve accuracy.

• 01Native multivariate and univariate forecasting capabilities
• 02Support for past-only and past-future (dynamic) covariates
• 03Non-autoregressive decoding for single-pass, low-latency inference
• 04Alternating attention architecture for capturing cross-series correlations
• 05Probabilistic forecasting with 9-quantile output for uncertainty estimation