Google AI Releases TimesFM-3: 330M Parameter Forecasting Model

Google AI has officially released TimesFM-3, a zero-shot foundation model designed for multivariate time series forecasting featuring 330 million parameters. This new release addresses the previous limitations of earlier versions that were restricted strictly to univariate forecasting.

In real-world forecasting scenarios, variables rarely operate in isolation. Google uses ice cream sales as an example, where related product sales, foot traffic, weather, promotions, and holidays all actively influence the target metric. Previous iterations like TimesFM 2.5 were only capable of forecasting a single series from its own historical data.

TimesFM-3 is built on a decoder-only transformer backbone. Consecutive data points are grouped into patches of 32 steps and normalized per series to prevent drastically different scales from dominating. Target and past covariate tokens originate from a single patch, while past-future covariates leverage a lookahead trick to let the model perceive scheduled events before they occur.

For decoding, TimesFM-3 adopts Contiguous Patch Masking to reduce latency, compute costs, and compounding errors found in earlier patch-by-patch methods. Alternating attention layers fill every masked horizon patch simultaneously, providing each target with 9 quantiles ranging from the 10th to the 90th percentile at every step.