Root mean squared error
Root mean squared error, or RMSE, is simply the square root of the mean squared error. Taking the square root converts the metric from squared units back t...
Bringing Error Back to the Original Scale
Root mean squared error, or RMSE, is simply the square root of the mean squared error. Taking the square root converts the metric from squared units back to the original units of the target variable, making it directly interpretable. A house price model with RMSE of 20,000 dollars is typically off by around 20,000 dollars, which is a statement a stakeholder can immediately understand.
RMSE inherits MSE's sensitivity to large errors because squaring still occurs before averaging. The square root moderates the penalty slightly compared to raw MSE but does not eliminate it. A model with a few very large errors will show a higher RMSE than one with many small errors of the same total absolute magnitude.
The relationship between RMSE and MAE provides useful diagnostic information. When RMSE is much larger than MAE, it signals that the error distribution has a heavy tail — there are a few large errors pulling RMSE up while the typical error, captured by MAE, is much smaller. When RMSE and MAE are close together, errors are spread more uniformly without extreme outliers.
RMSE is the standard metric for comparing regression models in most domains including weather forecasting, demand prediction, and financial modeling. Competition platforms commonly use RMSE as the official leaderboard metric because it strikes a balance between interpretability and sensitivity to large errors.
A concrete example: a wind energy forecasting model predicts daily output in megawatts. Predictions are mostly accurate but occasionally far off during unusual weather. MAE is 4.2 megawatts but RMSE is 11.8 megawatts, revealing that while most daily predictions are close, a handful of days see large errors that greatly inflate RMSE.
Exam tip: when comparing two models, use both RMSE and MAE. If one model has lower RMSE but higher MAE than another, the first model is making fewer extreme errors, which may or may not be more important depending on the application.