title: Simons Forecast Explorer
emoji: 📈
colorFrom: blue
colorTo: gray
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
short_description: Simons forecasts with and without known-future covariates
models:
- abel-lab/Simons-25M
- abel-lab/Simons-60M
- abel-lab/Simons-178M
- abel-lab/Simons-297M
Simons Forecast Explorer
A forecast explorer for Simons, a family of four small, open time series foundation models for probabilistic forecasting from Abel AI Lab. Pick an example series, a model size, a cutoff and a horizon, and compare the forecast with what happened next. Covariate examples show the forecast with and without known-future covariates in one chart.
Every forecast on the page was computed ahead of time with the simons package and the released
weights, so the page answers at once: 4 example series, 4 model sizes, 3 horizons and 3 cutoffs
each, with and without covariates where the series has them. Model details and weights are on the
model cards (25M,
60M,
178M,
297M) and in the
Simons 1 collection.
How to use it
- Choose a series and a model size: the 60M model is the default.
- Pick the cutoff: the model reads the history before it (up to 16,384 steps for the 25M and 60M models, 32,768 for the 178M and 297M models).
- Pick the horizon: the number of steps to forecast.
- Read the result: the plot shows the median forecast with its 10-90% and 30-70% bands against the held-out actuals. The table gives the MASE and CRPS of the held-out part for Simons and for a naive forecast (seasonal naive for hourly data, the last value for daily data), lower is better.
- Known-future covariates: where a series has them, the forecasts with and without them are drawn in one chart, with the covariates plotted below.
Example data
The ERCOT, Mississippi and Mauna Loa series end in early November 2026. The default forecast of the
Mississippi and Mauna Loa series covers their latest values; ERCOT's covers November 5, one full day. The Rossmann series ends in
July 2015. ERCOT's covariate is EIA's day-ahead demand forecast: each
day's forecast is made the day before, so the model reads it through the end of the day (Central
time) on which the forecast starts and forecasts later steps without it. The page stores the plotted
part of each series in data/.
| Example | Source | Dates | License |
|---|---|---|---|
| ERCOT electricity demand, hourly, with EIA's day-ahead demand forecast as a known-future covariate | U.S. Energy Information Administration, Hourly Electric Grid Monitor (Form EIA-930), balancing authority ERCO, demand and day-ahead demand forecast in MWh | 2026-08-01 00:00 to 2026-11-07 23:00 UTC | Public domain (EIA) |
| Mississippi River at St. Louis, daily | U.S. Geological Survey, gauge 07010000, daily mean discharge in ft³/s; recent values are provisional | 2019-01-01 to 2026-11-07 | Public domain (USGS) |
| CO₂ at Mauna Loa, daily | NOAA Global Monitoring Laboratory, daily mean CO₂ at Mauna Loa in ppm; days without a measurement are empty, and measurements from December 2022 to July 2023 come from Maunakea | 2016-01-01 to 2026-11-07 | Public domain (U.S. Government work); cite as Dr. Xin Lan, NOAA/GML (gml.noaa.gov/ccgg/trends/) and Dr. Ralph Keeling, Scripps Institution of Oceanography (scrippsco2.ucsd.edu/) |
| Rossmann store sales, daily, with promotion days, opening days and school holidays as known-future covariates | Rossmann Store Sales (Kaggle competition), store 1, daily sales; the series comes from fev-bench (rossmann_1D) |
2013-01-01 to 2015-07-31 | Kaggle competition rules |
Scores
- MASE: the mean absolute error of the median forecast, divided by the mean absolute error of a naive forecast over the history: seasonal naive with period 24 (the value a day earlier) for hourly data, and the last value for daily data.
- CRPS: the mean weighted quantile loss over the 9 quantiles (0.1 to 0.9), divided by the mean absolute actual value.
Files
The page is plain HTML, CSS and JavaScript with no build step: index.html, style.css, app.js,
the forecasts in data/ and the logos and fonts in assets/. It loads Plotly.js from its CDN.
The files in data/ were computed with the simons package and the released weights.
License
The code of this Space is licensed under Apache-2.0 and the Simons weights under OpenMDW-1.1. The
example data keep the licenses listed above. The Inter font files are under the SIL Open Font
License (assets/fonts/OFL.txt).