Zero-Shot Time-Series Forecasting: What It Means and When It Works

Zero-shot forecasting means using a pretrained model on a series it has never seen, without updating its weights. It is the fastest way to get a forecast from a new dataset: paste the history, choose a horizon, and read the output.

It is also easy to misuse. A zero-shot model is a strong baseline, not a substitute for understanding your data. This guide explains what it can and cannot do, and how to use it without fooling yourself.

CSV forecast studio

Paste a series, get a probabilistic forecast.

One value per row, or a date column plus one or more numeric columns. Timestral returns a median forecast and a likely range. The demo runs on the public Hugging Face Space; the API is the paid product.

1 · Paste or upload a CSV2 · Pick a horizon3 · Download the 21 quantiles

Upload CSV

Steps

Airline passengersMonthly salesWebsite visitsElectricity load

month,passengers 1949-01-01,112 1949-02-01,118 1949-03-01,132 1949-04-01,129 1949-05-01,121 1949-06-01,135 1949-07-01,148 1949-08-01,148 1949-09-01,136 1949-10-01,119 1949-11-01,104 1949-12-01,118 1950-01-01,115 1950-02-01,126 1950-03-01,141 1950-04-01,135 1950-05-01,125 1950-06-01,149 1950-07-01,170 1950-08-01,170 1950-09-01,158 1950-10-01,133 1950-11-01,114 1950-12-01,140 1951-01-01,145 1951-02-01,150 1951-03-01,178 1951-04-01,163 1951-05-01,172 1951-06-01,178 1951-07-01,199 1951-08-01,199 1951-09-01,184 1951-10-01,162 1951-11-01,146 1951-12-01,166

↗Run forecast1 credit per API call · the web demo is free while the public Space is available

How zero-shot forecasting works

A foundation model is trained on a large corpus of time series. It learns patterns that transfer across domains: trend, seasonality, level shifts, volatility clustering, missing-data behavior. At inference time, you provide a history and a horizon; the model conditions on the history and produces a forecast.

No gradient updates, no hyperparameter search, no training set. That is the whole appeal.

What it is good at

• Fast baselines. You can forecast a new series in minutes, not weeks.
• Many series. Score hundreds of SKUs, sensors or regions with the same model.
• Probabilistic output. Timestral returns 21 quantiles, so you get uncertainty without building a separate model.
• Cold start. A new product or a new region with a short history can still get a forecast.

Where it fails

1. Structural breaks

If the world changed in a way the history has never seen — a new competitor, a regulation, a pandemic — the model will extrapolate the old pattern. A zero-shot model has no causal knowledge. You need to encode the change as a covariate or override the forecast.

2. Promotions and interventions

A promotion is an intervention, not a pattern. If the history contains promotions, the model may learn an average effect, but it will not know the effect of a new promotion. Add a promotion flag as a covariate or model the lift separately.

3. Very short series

Most foundation models need enough history to infer a pattern. The studio requires at least 8 rows; a longer history usually produces a better forecast. If you have 3 points, use a human judgment or a simple rule, not a foundation model.

4. Series with no shared structure

Zero-shot transfer works when the new series resembles the training distribution. A highly specialized series with a unique data-generating process may need a custom model.

5. Causal questions

“What happens if I cut the price by 10%?” is a causal question. A forecasting model predicts the continuation of a pattern; it does not answer interventions.

How to set up a fair baseline

1. Hold out the last 20–30% of the series.
2. Forecast the holdout with the same horizon you will use in production.
3. Score the median with MAE/RMSE or MASE.
4. Score the interval with coverage: the actual should fall inside the 80% band about 80% of the time.
5. Compare with seasonal naive and at least one classical model.
6. Record latency and cost for the same batch.

If the zero-shot model does not beat seasonal naive, do not deploy it just because it is a foundation model. If it does, you have a baseline worth building on.

When to fine-tune

Fine-tuning makes sense when:

• You have a large, clean history for the target domain.
• The zero-shot model is close but consistently biased.
• You need to incorporate domain-specific covariates or constraints.
• The cost of a custom model is justified by the forecast value.

Mistral positions Timestral Forge for fine-tuning on proprietary data. Start zero-shot, measure the gap, then decide.

A practical workflow

1. Paste or upload a series.
2. Run a zero-shot forecast at the horizon you need.
3. Check the median and the 80% band against the last known values.
4. Backtest on a holdout.
5. If the error is acceptable, deploy; if not, add covariates or fine-tune.
6. Re-run the backtest when the data-generating process changes.

What to read next

• What is Timestral? — the model family.
• Timestral vs Chronos-2 — a testable comparison.
• How to evaluate forecasts — metrics and backtesting.
• API pricing — one credit per forecast call.

FAQ

Does zero-shot mean no data is needed?

No. You need a history for the series you want to forecast. Zero-shot means no training, not no input.

Can I fine-tune Timestral on my data?

Mistral positions Timestral Forge for fine-tuning. For many short-horizon problems, the zero-shot model is enough; measure the gap first.

Is zero-shot better than a classical model?

Not always. On short, stable series, seasonal naive or exponential smoothing can be hard to beat. Always include a classical baseline.