Google's TimesFM-3 Looks Beyond the Curve
Google Research has introduced TimesFM-3, a forecasting model built to predict future values in time series — the ordered streams of numbers that show up as daily sales figures, traffic counts or sensor readings — while also weighing the context that surrounds those numbers. The release marks another step in the push to turn forecasting into a general-purpose capability rather than a bespoke engineering project rebuilt for every dataset.
The central premise is that real forecasts rarely hinge on a single variable. Google illustrates the point with a retail chain trying to anticipate ice cream sales. A model that looks only at past ice cream purchases will miss a great deal: sales of related items such as waffle cones and syrup, historical foot traffic near the store, weather conditions, planned discount campaigns and holidays all shape demand. TimesFM-3 is designed to fold those signals into one prediction instead of treating each as a separate problem.
Patches, Normalization and a Two-Directional Transformer
Architecturally, TimesFM-3 remains in the Transformer family, the same base design used by its predecessors. But it makes deliberate choices about how time series data enters the network. Rather than feeding the model one data point at a time, it groups 32 consecutive points into a single patch. That patch becomes the unit the model reasons over, which shortens the sequence and gives the network a broader view of local shape.
Google also normalizes each series to a common scale. That matters because time series arrive in wildly different magnitudes — a handful of units sold per day versus millions of readings from an industrial sensor — and comparing them directly would be meaningless. Scaling each series to a shared baseline lets the model treat patterns from very different domains as commensurable.
Processing then runs in two alternating directions. Along the time axis, the model hunts for patterns inside a single series, and it draws only on past values so that future information cannot leak into the prediction. Across series, it compares every variable at a given moment and learns how they relate. That cross-series view is what allows the model to pick up on effects such as a discount on one product shifting sales of another — a relationship a single-series model would never observe.
Google's illustration of the input setup color-codes the data: a blue target series, purple variables known only from history, and green tokens for known future events whose representation already includes the upcoming patches.
Three Flavors of Extra Signal
TimesFM-3 accepts three distinct types of supplementary data, according to Google.
- Covariates it predicts jointly. The model can forecast several related variables at once, such as demand across different ice cream flavors, rather than treating each one in isolation.
- Historically known variables. Factors such as past foot traffic are available up to the present but not beyond it, and the model incorporates them as context.
- Known future events. Planned promotions, discount schedules and weather forecasts are events whose timing is already known, so the model can use them as forward-looking inputs rather than inferring them from history alone.
That third category is where TimesFM-3 diverges most sharply from classical statistical forecasting. A promotion calendar is not a pattern waiting to be discovered — it is a fact about the future. This model is engineered to accept such facts as input.
One Shot Instead of Step by Step
Earlier versions of the model generated predictions one block at a time. Each new block was built on the previous prediction, an approach Google describes as slow, compute-heavy and vulnerable to compounding error: a small mistake early in the sequence could distort everything that followed. This is a familiar failure mode in autoregressive forecasting, where a model effectively feeds on its own output.
TimesFM-3 abandons that loop. It assigns blank placeholders to every future time step and completes them in one sweep. Google frames this as one-shot forecasting, and the practical difference shows up in the ice cream scenario. A model that knows only past sales will simply continue the established weekly rhythm, blind to a promotion that has not happened yet. Once TimesFM-3 receives the discount schedule, it can adjust the forecast in a single pass rather than walking forward increment by increment.
Nine Values per Step: Uncertainty Built In
Instead of emitting a single point estimate for each time step, TimesFM-3 outputs nine values per step. Those values are intended to capture the range and uncertainty of the prediction — a forecast that conveys not just what is likely to happen, but how confident the model is. For planning purposes, that distinction is often more useful than a bare number, because it lets downstream systems reason about risk, buffer inventory or spare capacity.
Trained at Trillion-Point Scale, Deployed Zero-Shot
The model carries 330 million parameters and was trained on a mix of real and synthetic time series totaling more than one trillion data points, according to Google. Like its predecessors, it operates zero-shot: it requires no additional training for a new task. That property is central to the pitch. Rather than assembling a labeled dataset and fitting a model for every new forecasting problem, a practitioner can point TimesFM-3 at a series and its associated signals and expect usable output.
Zero-shot performance is also what makes a foundation-model approach to forecasting attractive in the first place. Forecasting problems are everywhere — inventory, energy load, staffing, logistics, capacity planning — but each historically demanded its own pipeline, its own features and its own validation regime. A single model that generalizes across them changes who can build a forecasting system and how quickly.
The breadth of the training data matters here as much as the parameter count. By mixing synthetic series with real ones, Google can expose the model to a wider variety of shapes, seasonal rhythms and shock patterns than any single domain could supply. That diversity is the raw material for generalization, and it is what the zero-shot claim ultimately rests on.
The Broader Bet on Foundation Models for Numbers
TimesFM-3 sits at the intersection of two trends: the scaling of foundation models beyond text and images, and the long-standing demand for better operational forecasting. Google's framing is that context is what separates a naive extrapolation from a useful prediction, and the architecture is a direct response to that claim — patches for efficiency, normalization for comparability, attention across both time and variables for relationships, and a one-shot decoder for coherence.
Whether that combination holds up across messy real-world data remains the open question, as it does for any zero-shot system. But the direction is clear: forecasting is being treated less as a statistical chore and more as a general reasoning task, one in which sales data, weather and a discount calendar all belong in the same input. For retailers running promotions, utilities balancing load and operators planning capacity, the appeal is straightforward — the model is asked to anticipate events that are already on the calendar, not merely to extend last year's line.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
