Timestral vs Chronos-2 vs Statistical Baselines: A Testable Comparison

Model comparisons are usually written by someone selling one of the models. The useful question is not “which model wins a leaderboard” but “which model produces a forecast I can act on, on my series, at a cost and latency I can afford.”

This guide compares Timestral with Chronos-2 and classical baselines on the axes that decide a production choice, and gives you a backtest you can run 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

The short version

• Timestral is a family of four open models (5M, 28M, 123M, 209M) for zero-shot probabilistic forecasting. Mistral reports that the 209M model leads GIFT-Eval, TIME and fev-bench among pretrained zero-shot models, and that the 28M model beats Chronos-2 on GIFT-Eval and TIME. Those are vendor claims; test them on your data.
• Chronos-2 is a strong open forecasting model with a different architecture and training recipe. It is a serious baseline, especially for univariate series.
• Classical baselines (seasonal naive, exponential smoothing, ARIMA) are cheap, interpretable and surprisingly hard to beat on short, stable series. Always include at least one.

The five axes that matter

1. Zero-shot quality

Does the model produce a usable forecast on a new series without training? Test on a holdout window from your own data. Do not compare on a single series; use at least 20 series or 20 rolling windows.

2. Probabilistic output

Does the model return a distribution or only a point? Timestral returns 21 quantiles. Chronos-2 also produces probabilistic forecasts. A point-only baseline can still be useful, but it cannot support risk decisions without an extra model.

3. Multivariate support

Can the model use several related series together? Timestral supports multiple numeric columns (up to 8 in the studio). If your problem is genuinely multivariate, a univariate model that ignores the other series may leave accuracy on the table.

4. Context length

How much history can the model use? Timestral supports up to 32,768 time steps. Long context helps with slow cycles and long-range dependencies, but it also increases compute and can add noise. Test a capped history against the full history.

5. Cost and latency

The cheapest model is the one that clears your accuracy bar. Timestral is designed for CPU inference; Mistral reports roughly 10 ms for the 5M model and 140 ms for the 209M model on a standard eight-thread CPU for a one-day-ahead forecast from three weeks of 15-minute data. Chronos-2 and classical models have different profiles. Measure on your hardware.

A backtest you can run

1. Pick 20 series that look like your production data. Include stable, seasonal, sparse and noisy series.
2. Create rolling windows. For each series, use the first 70% as history and the next 30% as test. Roll forward in steps equal to your forecast horizon.
3. Forecast with every candidate. Same history, same horizon, same covariates. Do not give one model a better prompt or a longer context.
4. Score the median. Use MAE or RMSE for scale-dependent series, and MASE or sMAPE for scale-free comparisons.
5. Score the interval. Measure the percentage of actuals inside the 80% band. It should be close to 80%.
6. Measure latency and cost. Record wall-clock time and the price per 1,000 forecasts.
7. Pick the simplest model that clears the bar. If seasonal naive is within 2% of the foundation model, use seasonal naive and spend the complexity budget elsewhere.

What the benchmarks do and do not tell you

GIFT-Eval, TIME and fev-bench are useful because they cover many datasets and horizons. They are not your data. A model that leads a benchmark can still lose on a series with a strong weekly pattern, a promotion, or a structural break.

Also, benchmark leadership among “pretrained zero-shot models” is a narrower claim than “best forecasting model.” A fine-tuned model, a classical model with the right features, or a human planner can beat any zero-shot model on a specific series.

When Timestral is a good fit

• You need a probabilistic forecast quickly, without training.
• You have several related series and want to forecast them together.
• You want CPU inference and predictable latency.
• You need a long context for slow cycles or long histories.
• You are building a baseline before investing in a custom model.

When another approach is better

• Your series is short and stable: seasonal naive or exponential smoothing may be enough.
• You have strong causal features: a supervised model with those features may beat a zero-shot foundation model.
• You need explainability for every coefficient: a classical model is easier to defend.
• You need sub-millisecond latency at massive scale: a small custom model may be cheaper.

What to read next

• What is Timestral? — the model family.
• How to evaluate forecasts — the metrics and backtest design.
• Probabilistic forecasting and the 21 quantiles — reading the band.
• API pricing — one credit per forecast call.

FAQ

Is Timestral better than Chronos-2?

Mistral reports that Timestral-28M beats Chronos-2 on GIFT-Eval and TIME, and that Timestral-209M leads several zero-shot benchmarks. The only reliable answer is your own backtest on your series.

Can I use both?

Yes. Run both as candidates in the backtest and pick per series or per horizon. A model ensemble is also an option, but measure whether it beats the best single model after costs.

Do I need a GPU for Chronos-2 or Timestral?

Not necessarily. Both can run on CPU at modest volumes. Measure latency on your hardware; the model size and context length matter more than the brand.