WEB Signal 400
Google's TimesFM-3 adds multivariate forecasting, covariates and one-pass horizon prediction to 330M-parameter model
↗
TimesFM-3 is Google's 330M-parameter time-series foundation model that forecasts multiple related series jointly, incorporates past and future-known covariates, and generates an entire forecast horizon in a single forward pass.
The shift from univariate to multivariate forecasting with native covariate support means teams can model related series together, such as product categories sharing a promotional calendar, without separate model calls. One-pass horizon generation via Contiguous Patch Masking removes the patch-by-patch decoding loop that added latency and let early-patch errors cascade into later predictions in earlier versions.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
TimesFM-3 moves beyond the univariate design used through version 2.5, jointly modeling multiple target channels with past-only and past-and-future covariates.
Contiguous Patch Masking generates the full forecast horizon simultaneously instead of decoding patch by patch, reducing latency and preventing early-patch errors from cascading.
The model ranked first among pretrained foundation models on GIFT-Eval, fev-bench, and the TIME leaderboard for both point and probabilistic forecasting metrics.
THE CLUSTER