We can leave supervised learning to one side for the length of this piece. It earns that exclusion because it relies on precisely what a zero-shot model removes: a labelled training set built for the task in hand, and rebuilt whenever the task shifts.
Picture a retail manager forecasting ice cream sales for one shop in the month ahead. The conventional route reads that shop's own history, projects the weekly shape forward, and stops there. It cannot know that a heatwave is forecast, or that a promotion lands in the second week of July. Capturing those drivers normally means collecting weather and promotion history, labelling it, joining it to sales, and training a model that serves one store.
Zero-shot forecasting removes the labelling and the training job. The model returns a forecast for a series it was never optimised on, with no gradient updates at inference, as Emergent Mind's survey of the capability sets out.
The Foundation Model Paradigm
Foundation is the word carrying the weight. A conventional model is built for a single job, whether that is mortality rates or stock levels, and it stays there. A foundation model is pre-trained across a wide corpus of time series and learns the recurring shapes: trend, seasonality, level shifts, and the texture of the noise around them.
That prior is the product. Hand the model a new sequence at inference and it applies patterns it already holds, rather than patterns it has just been taught on your data. The gain shows up most clearly where history is short. Predict.ai's guide to the technique makes the cold-start case plainly: a new region that resembles old ones, a short series, or a first pass on a Tuesday afternoon when the alternative is a blank cell. Its warning is just as useful, because a long and idiosyncratic series with a private promotion calendar can lose to last week's number.
Moving from Univariate to Multivariate Reality
Early foundation models were univariate. They read one series and forecast that series. Most real problems are not shaped that way, and the ice cream example shows why: cone and syrup sales, foot traffic, and a known promotion calendar all move the target.
TimesFM-3 is the generation that changes this. Google Research states that it predicts coevolving series jointly and captures relationships that "improve overall accuracy without requiring task-specific fine-tuning". It accepts two broad kinds of input:
- Several related targets at once, such as different ice cream brands forecast together.
- Covariates, whether they are known only historically, such as past foot traffic, or known into the future, such as a promotion schedule or a weather forecast.
The second kind is where the design earns its keep. Target series and past-only covariates enter as ordinary patches. Covariates known ahead of time get a lookahead treatment, where each token carries the current patch together with upcoming patches, so the model can see a promotion before it happens. Attention then alternates: causal attention runs along each series in time, so a token cannot read its own future, while variate attention runs across series at a fixed step, letting cross-series correlations surface.
Learning to use a covariate without training on the dataset is the hard part, and it is an active research problem in its own right. The COSMIC work on zero-shot forecasting with covariates frames it as in-context learning, because the model has to infer how a covariate relates to the target from the context window alone. Nothing is fitted to the task, so the relationship has to be read on the spot. That is precisely the claim Google is making when it says a promotion calendar will move an ice cream forecast without anyone retraining a store-level model.
Efficiency through Non-Autoregressive Decoding
A significant bottleneck in previous models was autoregressive generation, where the model predicted one step, fed that prediction back into itself, and predicted the next. That creates latency and lets errors compound along the horizon.
TimesFM-3 fills the horizon in one forward pass. It places masked placeholders for the unknown future values while keeping known future signals, such as the promotion schedule, visible. It then predicts the whole window at once and returns a set of quantiles for each target at each step, so uncertainty arrives with the forecast rather than after it. On a promotion day, the bump appears without the model iterating day by day to find it.
The Evidence-Access Framework
The reported benchmark averages put TimesFM-3 at the top among pre-trained foundation models on both point accuracy and probabilistic quality across Gift-Eval, FEV-Bench and Time, and it holds that position even when run in univariate mode, with no covariates and no cross-series information.
Scores alone cannot carry the claim much further, because "zero-shot" by itself says very little. It describes a training condition, not an evidence condition, and the difference matters the moment two systems are compared. The Pith audit framework sets out the problem: a frozen language model prompted with serialised values, a time-series model pre-trained on forecasting corpora, and a retrieval-augmented system can all satisfy the no-update condition while drawing on fundamentally different evidence. Classify a model by where its evidence comes from, and a leaderboard row becomes a claim someone can audit rather than a rank they have to trust.
Implementation and Risks
For software teams the appeal is immediate: move from data to forecast without the overhead of a bespoke training pipeline. The risk is treating the output as more than it is. In clinical settings, generative models behave as simulators rather than oracles, producing many plausible futures and counting how often an event appears among them. Stanford HAI has argued that those simulated frequencies are not real-world probabilities until a model is calibrated and audited, and that acting on them can push a team into overtreating low-risk patients or missing high-risk ones.
Evidence from outside a vendor's own benchmarks is worth reading before any commitment. An independent study of mortality-rate forecasting across fifty countries and 111 age groups found a mixed picture: CHRONOS was competitive at short horizons and beat ARIMA and the Lee-Carter model, the earlier TimesFM consistently underperformed, and fine-tuning improved CHRONOS's long-term accuracy. A random forest trained on the mortality data itself won overall. That study is a reminder that a universal prior is not a substitute for a domain-fit check.
The working practice, then, is a tournament rather than a favourite. Run the zero-shot model against a seasonal naive baseline and against whatever you already have, on a holdout that resembles the future you care about, and keep it only if the prior adds something measurable. That backtest is the forecasting equivalent of the validation layer that How to Evaluate CI/CD Pipelines treats as non-negotiable, because automation is only worth having when something independent checks its output. It is also an abstraction bargain of the kind Making Low-code Platforms Work examines elsewhere: fine control over the model is exchanged for speed of first deployment, and the trade holds only while the defaults hold.
TimesFM-3 is available now on GitHub and Hugging Face, with BigQuery integration following, so running that test costs an afternoon rather than a quarter.
Sources
- Zero-Shot Forecasting Capability: Surveys zero-shot forecasting, the architecture families used, and the transfer mechanisms behind them.
- Zero-shot forecasting: Explains the cold-start case for foundation models and where fine-tuning later pays off.
- Zero-Shot Time Series Forecasting with Covariates via In-Context Learning: Shows how a zero-shot model can learn covariate-target relationships from context alone.
- Tracing the Evidence Behind Zero-Shot Time-Series Forecasting: Proposes a source-first taxonomy that classifies zero-shot systems by their primary evidence source.
- Why 'Zero-Shot' Clinical Predictions Are Risky: Distinguishes generative simulation from calibrated prediction in high-stakes settings.
- Zero-Shot Forecasting Mortality Rates: A Global Study: Benchmarks zero-shot and traditional forecasters across mortality data for 50 countries.
improve overall accuracy without requiring task-specific fine-tuning
Source: TimesFM-3: A zero-shot foundation model for multivariate forecasting