RMSE Between Two Lists

Root Mean Squared Error (RMSE)

The RMSE tells you how far off your predictions tend to be from the observed values: RMSE = √(Σ(yᵢ − ŷᵢ)²/n). One handy property is that it comes out in the same units as the target variable, so the number actually means something you can reason about. Squaring each error means the metric punishes large misses much harder than the MAE (Mean Absolute Error) does, since MAE stays linear. Take y = 3, 5, 2.5, 7, 9 and ŷ = 2.5, 5, 4, 8, 8. The squared errors are 0.25, 0, 2.25, 1, 1, summing to 4.5, so RMSE = √(4.5/5) ≈ 0.949. There's also a tie to the coefficient of determination, R² = 1 − SSE/SST. You'll see RMSE as the default regression metric in sklearn.metrics and across plenty of Kaggle competitions.

Applications

It shows up when you pick between ML models (regression, neural nets, gradient boosting), when you check climate and weather models against actual measurements, in demand and sales forecasting on time series, in GPS positioning error, in scientific calibration, and in Kaggle / DrivenData competitions.

FAQ

RMSE or MAE? Reach for RMSE when a big miss really hurts and should weigh more (think safety, finance). Stick with MAE when every error counts the same and you'd rather not let outliers dominate.

Is the RMSE always positive? Yes. It's the square root of a sum of squares, so it can't go below zero. A value of 0 means your predictions hit perfectly, and the bigger it grows, the worse the model is doing.

Can I compare RMSEs across different problems? Not directly, because the value scales with y. If you need a comparison, normalize it first: NRMSE = RMSE / (ymax − ymin) or RMSE / mean(y).