Time-series playground
A clearer view of what’s next.
Real public data: NYC subway ridership, Wikipedia traffic, and Mauna Loa CO₂. Or upload your own. Fifteen models compete on a chronological holdout, and the winner forecasts forward, all in your browser.
Loading…
- Observed
- 7-day average
- Backtest
- Forecast
- Heuristic range
- Holdout window
How fast the level, trend, and seasonal pattern react to new data. Higher values chase recent changes; lower values smooth them out.
Model leaderboard
Click a model to forecast with it instead of the auto pick.
Read the signal
How the evaluation works
Each model trains on the first 80% of observations and predicts the final 20% without seeing it. The lowest mean absolute error wins, then that model is refit on the full series. The shaded range is ±1.96 × holdout RMSE, widened with distance: a heuristic, not a calibrated interval. The bundled series are real public data, refreshed with a script in the repo; uploads never leave your browser. None of this is financial or operational advice.
Bring your own data. A CSV with date,value columns, 21 to 5,000 evenly spaced rows, ISO dates, under 500 KB.