Time Series Forecasting Lab
Upload your time series, compare forecasting models with chronological backtesting, and generate data-driven predictions. Everything runs locally in your browser.
Forecast overview
Historical & forecast
Prediction intervals require sufficient independent historical forecast errors.
Model leaderboard
Click a row to inspect its forecast| Model | MAE ↓ | RMSE ↓ | sMAPE ↓ | Bias |
|---|---|---|---|---|
| Run an analysis to compare models. | ||||
Backtesting diagnostics
Actual vs predicted
Mean absolute error by forecast horizon
The leaderboard uses equally weighted forecast errors across all backtest folds. Positive bias means overforecasting.
Learn the concepts
Why chronological backtesting?
Random train/test splitting leaks future information into time-series evaluation. Expanding-window backtesting fits models on the past and evaluates genuinely future periods.
What do MAE, RMSE, and sMAPE measure?
MAE is the average absolute forecast error. RMSE gives greater weight to large errors. sMAPE scales errors by the magnitude of actual and predicted values; it can behave poorly near zero.
Why is the seasonal model unavailable?
Seasonal Naive requires at least two full seasonal cycles in every training fold. The candidate seasonal periods are 7 for daily, 12 for monthly, and 52 for weekly data.
How reliable are prediction intervals?
Intervals use historical out-of-sample forecast errors. They are displayed only when at least 20 non-overlapping backtest folds provide an error for each horizon. These empirical intervals are not guaranteed to achieve 90% future coverage.
Is the best model guaranteed to win in the future?
No. Backtest results are estimates based on historical data. Selecting the winner using these same results can make the winner's reported performance optimistic. Forecasts assume future behavior remains sufficiently similar to the past.