Timestral

Use case

Energy price forecasting with a range for volatile hours

Forecast hourly energy prices from a CSV, read the daily shape and the 80% band, and use the range to plan procurement and battery dispatch.

Run the samplePut this in production

Energy prices are volatile, and a point forecast can be worse than no forecast if it hides the tail. The useful output is a distribution: a median path and a range that widens around the volatile hours.

The sample is hourly day-ahead price in USD/MWh with a morning and evening peak. Keep the market, the currency and the unit explicit. A price series from a different market is not a covariate; it is a different process.

The sample series

The studio below is preloaded with this hourly price (usd/mwh) series and a 24-step horizon. It is an illustrative example, not a real company’s data. Replace it with your own CSV to run the same workflow.

Show the CSV
time,price_usd_mwh
2026-05-01 00:00,28.4
2026-05-01 01:00,26.1
2026-05-01 02:00,24.8
2026-05-01 03:00,24.2
2026-05-01 04:00,25.6
2026-05-01 05:00,29.8
2026-05-01 06:00,38.2
2026-05-01 07:00,46.5
2026-05-01 08:00,51.1
2026-05-01 09:00,49.8
2026-05-01 10:00,47.2
2026-05-01 11:00,45.6
2026-05-01 12:00,44.1
2026-05-01 13:00,43.8
2026-05-01 14:00,45.2
2026-05-01 15:00,48.9
2026-05-01 16:00,54.6
2026-05-01 17:00,59.8
2026-05-01 18:00,62.4
2026-05-01 19:00,58.1
2026-05-01 20:00,51.2
2026-05-01 21:00,44.6
2026-05-01 22:00,37.9
2026-05-01 23:00,32.4
2026-05-02 00:00,29.1
2026-05-02 01:00,26.8
2026-05-02 02:00,25.2
2026-05-02 03:00,24.6
2026-05-02 04:00,25.9
2026-05-02 05:00,30.4
2026-05-02 06:00,39.1
2026-05-02 07:00,47.2
2026-05-02 08:00,52.0
2026-05-02 09:00,50.4
2026-05-02 10:00,47.9
2026-05-02 11:00,46.1
2026-05-02 12:00,44.8
2026-05-02 13:00,44.2
2026-05-02 14:00,45.8
2026-05-02 15:00,49.6
2026-05-02 16:00,55.3
2026-05-02 17:00,60.4
2026-05-02 18:00,63.1
2026-05-02 19:00,58.7
2026-05-02 20:00,51.8
2026-05-02 21:00,45.1
2026-05-02 22:00,38.4
2026-05-02 23:00,33.0

CSV forecast studio

Forecast your energy price forecasting series.

This studio is preloaded with the hourly price (usd/mwh) example above. Replace it with a CSV of your own history, choose a horizon, and run the forecast. The public demo is free while the Hugging Face Space is available; the API is the paid product.

1 · Paste or upload a CSV2 · Pick a horizon3 · Download the 21 quantiles
1 credit per API call · the web demo is free while the public Space is available

What to watch in the output

Use the band for dispatch thresholds

A battery dispatch rule that uses only the median will miss the high-price hours that create value. Compare the high band with your marginal cost, not just the median.

Keep the market and unit explicit

USD/MWh, EUR/MWh and cents/kWh are different scales. A unit mismatch can turn a forecast into a plausible-looking error.

Expect the band to widen at peaks

If the range is narrow during the evening peak, either the history is unusually stable or the model is overconfident. Check coverage in backtests before trusting it.

Common mistakes

  • —Mixing day-ahead and real-time prices. They are different series with different volatility.
  • —Ignoring negative prices. If the market can go negative, keep the zero-clipping option off.
  • —Using a price forecast as a trading signal without a backtest. A statistical forecast is not an edge.

How to put it to work

  1. 1. Start with the sample. Run the preloaded series once so you see the output shape before you debug your own CSV.
  2. 2. Replace it with your history. Keep one timestamp column and one or more numeric columns. Mark outages and stockouts as empty cells, not zeros.
  3. 3. Backtest before you plan. Hold out the last few periods, compare the forecast with what happened, and check whether the actuals land inside the 80% band.
  4. 4. Move it into production. Use the REST API for a nightly job or an internal tool. One credit per successful call; failed calls are refunded.

FAQ

Can the model handle negative prices?

Yes. Keep the “keep forecasts at or above zero” option off. Negative prices are a real market outcome, not a data error.

Can I include load or weather as a second column?

You can include related numeric columns, but the model learns correlation, not market physics. Validate any covariate with a backtest before using it for dispatch.

What horizon is useful for battery dispatch?

Day-ahead dispatch typically uses 24 hourly steps. The studio caps the UI at 36; the API supports longer horizons up to 1,024.

Need this as an API?

The same forecast is available as a metered REST endpoint with JSON and a downloadable quantile CSV.

See the API

Related reading