RMSE (Root Mean Square Error)
Average prediction error that penalizes big misses more than small ones. Lower is better.
RMSE measures prediction error with a deliberate bias toward large misses. Square each error, average them, then take the square root. The squaring step amplifies outliers: one terrible prediction drags your score down much more than several small misses would. This makes RMSE higher than MAE whenever your errors are inconsistent. If RMSE ≫ MAE, you have outlier predictions worth investigating. If RMSE ≈ MAE, your errors are uniform across all predictions.
How It Works
RMSE = √(mean((predicted − actual)²)). Units match your target variable. Compare RMSE to MAE: if they’re similar, errors are consistent; if RMSE is much larger, you have outlier failures.
Example
During Oil v16 forward testing, RMSE revealed something MAPE missed: while average percentage error was 2.26%, a handful of sessions with geopolitical spikes produced outsized errors that dragged RMSE up. This led directly to recalibrating Monte Carlo disruption parameters. Tracked in Oil v16 Sell Model.