TimesFM-3: Google's 330M-Parameter Multivariate Forecaster

TimesFM-3, released by Google Research on August 31, is the biggest shift in open time-series forecasting this year: a 330M-parameter model that does native multivariate forecasting in a single forward pass, zero-shot, with weights live on Hugging Face. Every previous TimesFM generation was strictly univariate — one series at a time, one patch at a time. TimesFM-3 ends both limits at once.

What shipped in TimesFM-3

Google pre-trained TimesFM-3 on a corpus of 1+ trillion time points (real-world and synthetic) and reworked the architecture around the LLM playbook — patch, mask, decode everything at once:

• Multiple targets, one pass: jointly forecasts coevolving series with both point and 9 quantile (10th–90th percentile) outputs per horizon step.
• Covariate support: past covariates (e.g. historical foot traffic) and past-future covariates (known future events like promotions or weather forecasts) — a lookahead token strategy lets the model peek at upcoming known signals.
• Single-pass decode: Contiguous Patch Masking fills the entire horizon at once — no autoregressive loop, no error accumulation, lower latency.
• Alternating attention on a 2D grid: causal temporal attention across time plus full variate attention across series, so one series' signal can lift another's forecast.
• Top-ranked on all three public benchmarks — Gift-Eval, FEV-Bench, TIME — beating Chronos-2 and the Toto 2.0 family on point and probabilistic metrics, even in plain univariate mode.
• Open weights: google/timesfm-3.0-pytorch on Hugging Face, code on GitHub, BigQuery integration landing in "the coming weeks."

Why TimesFM-3 matters

The demo in Google's post is the tell: feed a retailer's promotion calendar in as a past-future covariate and the forecast anticipates a ~20% sales lift on each promo day — no fine-tuning, no bespoke training. That's the multivariate unlock: forecasters can finally condition on the things they already know are going to happen.

It also speaks directly to the forecast-collapse failure Dispatch covered in August, where univariate TSFMs went flat ranking low-signal targets like equity returns. Variate attention lets series borrow signal from each other instead of being predicted in isolation. Whether that fixes cross-sectional collapse on financial returns is the test worth running — the weights are open, so anyone can.

Verdict: this is the model to beat, and it's the first TSFM that behaves like a real foundation model rather than a fine-tune target.