TimesFM-3: Google's Native Multivariate Forecasting with One-Pass Decoding

Time series forecasting has long been a target for deep learning to surpass traditional statistical methods and hand-tuned pipelines. Since 2024, Google's TimesFM series has treated forecasting as a language modeling problem, leveraging the proven scaling and pre-training recipe of text models.

From Univariate to Multivariate: Why It Matters

TimesFM and TimesFM-2.5 (released September 2025) were strictly univariate models — each forecast used only the history of a single series. Real-world tasks, however, are often multivariate: a signal's future depends on other signals. For example, forecasting ice-cream sales for a retail chain requires not just past sales but also covariates such as cone sales, foot traffic, weather forecasts, and planned promotions. Ignoring these covariates degrades accuracy.

TimesFM-3 addresses this by accepting multiple target series, historical-only covariates, and past-future (dynamic) covariates whose future values are known in advance (e.g., promotion calendars, weather forecasts). The model ingests all this information and produces a joint forecast for all targets in one forward pass.

What Is TimesFM-3

330 million parameters Pre-trained on a corpus of real-world and synthetic time series totaling over 1 trillion time points Supports zero-shot generalization — no task-specific fine-tuning required Outputs both point forecasts and quantile forecasts (9 quantiles from the 10th to the 90th percentile) for probabilistic uncertainty estimation

"Native multivariate" means cross-series relationships are learned directly during pre-training and inference, not by stitching together independent univariate runs.

Internal Mechanism: How TimesFM-3 Works

TimesFM-3 retains the decoder-only Transformer backbone of its predecessors. The main changes are in input representation and attention, enabling parallel processing of multiple sequences.

Patching & Tokenization

Time series are still processed in patches. TimesFM-3 groups consecutive data points into patches of 32 time steps, so the model attends to patches rather than individual steps, reducing sequence length and compute.

Each patch is normalized per time series to handle scale differences — one series may be in single digits while another reaches millions.

Target series and historical covariates map one token per patch. Past-future covariates are handled differently: because their future values are known, each token concatenates the current patch with the future patch, letting the model read upcoming known signals (e.g., "day 14 has a promotion") as conditioning information.

Alternating Attention Architecture

After tokenization, data enters the Transformer as a 2D grid. Attention alternates between two modes across layers:

Causal temporal attention (horizontal): Tokens attend only within the same series along the time dimension, strictly respecting causal order (a token sees only past tokens). This prevents leakage and preserves each series' temporal structure.

Full variate attention (vertical): At the same time step, tokens can attend to all series in the dataset. Cross-series relationships enter here — e.g., whether a promotion on one product line lifts another, or whether foot traffic and ice-cream sales move together.

Alternating these two attention types lets the model jointly learn temporal patterns and cross-series dependencies in a single Transformer, without custom graph structures or separate encoders per series.

Non-Autoregressive Decoding: One Forward Pass

Previous TimesFM versions used autoregressive generation: predict one patch, feed it back, repeat. This works but latency grows with horizon, errors can accumulate, and compute cost rises.

TimesFM-3 replaces this with Contiguous Patch Masking.

Before forecasting, the model appends placeholder tokens for the entire prediction horizon after the observed context and masks the unknown portions. Target series and historical covariates keep their future patches masked; past-future covariates remain visible because their future values are known. A single forward pass then fills all masked patches at once, eliminating iterative steps and error accumulation. The output is a full probabilistic forecast — 9 quantiles per target per prediction step.

Concrete Example: Promotions and Ice-Cream Sales

Google Research illustrates the univariate vs. multivariate gap with a retail promotion scenario. Assume next month's promotion schedule is fixed. A univariate model sees only historical sales, extracts weekly seasonality, and projects a smooth curve that merely repeats the past — it cannot anticipate the sales lift on promotion days.

The univariate model does not know which dates have promotions, so its forecast stays flat; the typical promotion-driven uplift does not appear in the prediction.

TimesFM-3 receives the fixed promotion calendar as a past-future covariate. Having learned the promotion–sales relationship during pre-training, the model predicts roughly a 20% sales increase on promotion days when it sees the covariate signal approaching. The final forecast shows visible responses for each promotion window, while the univariate baseline follows the original trend. Over a full month, this gap compounds into materially more accurate revenue forecasts.

Such scenarios are not contrived: retail demand planning, electricity load forecasting with weather, patient monitoring with multiple vitals, and financial risk modeling with correlated asset prices all benefit from multivariate information.

Benchmarks: Where TimesFM-3 Stands

Google evaluated TimesFM-3 on three public forecasting benchmarks: Gift-Eval, FEV-Bench, and Time. Across both point and probabilistic forecasting, TimesFM-3 ranks first among pre-trained foundation models on all three benchmarks.

Key competitors included:

Chronos-2 — a multivariate-capable foundation model Toto 2.0 family — another recent model series in this space TimesFM-2.5 — Google's own previous generation

Results are reported as average rank across tasks (lower is better). TimesFM-3 was tested in two modes:

Univariate mode: No covariates, no cross-series information, each target handled independently. Even in this restricted setting, TimesFM-3 matches or beats competing models, showing the backbone's strong temporal modeling.

Full multivariate mode: With covariates and cross-series information enabled, TimesFM-3 achieves the best average rank on all three benchmarks for both point and probabilistic forecasting.

Non-autoregressive decoding also improves inference efficiency. Compared to patch-by-patch generation, TimesFM-3 is faster across prediction horizons — a practical difference when forecasting millions of SKUs for a retail chain or thousands of metrics in a cloud observability pipeline.

Why This Release Matters

Interest in time-series foundation models is growing, but moving research into production remains hard. From a practitioner's perspective, TimesFM-3 offers several critical advantages:

Zero-shot generalization. No fine-tuning on specific datasets. Teams without dedicated ML infrastructure can feed data and covariates directly to the model and obtain forecasts, drastically reducing upfront engineering cost.

Out-of-the-box probabilistic forecasting. The 9 quantiles provide a full uncertainty distribution, not just a point estimate. Inventory planning, for instance, needs stock-out probabilities — decisions that require probabilistic information.

Single forward pass. The non-autoregressive design avoids the latency and error accumulation of autoregressive forecasting, directly impacting usability for real-time or near-real-time tasks.

Upcoming BigQuery integration. Teams already on Google Cloud can run TimesFM-3 directly in BigQuery; TimesFM-2.5 is already accessible via the AI.FORECAST command. This lowers the adoption barrier further by removing the need for full ML engineering stacks.

Closing Thoughts

Competition among time-series foundation models is accelerating. TimesFM-3's shift from univariate to native multivariate forecasting, its move to non-autoregressive decoding, and its strong zero-shot performance across multiple benchmarks make it a compelling option for practitioners who need accurate forecasts without building custom pipelines from scratch.