The root mean squared error (RMSE) is the most widely reported single number for regression quality, and for one simple reason: it is the MSE brought back into the units you actually care about. You compute the average of the squared residuals and then take the square root, undoing the squaring so the final figure reads in the same units as the target. Low RMSE means tight predictions; high RMSE means the model is missing, often badly.
What is the RMSE?
RMSE takes every residual, squares it, averages the squares, and then takes the square root of that average. The squaring step means large residuals dominate the total, and the square-root step converts the result back from squared units into the target’s natural units. So the RMSE keeps the outlier-punishing behaviour of squared residual while remaining as easy to read as the MAE. That blend of sensitivity and readability is why it dominates leaderboards and reports.
The root mean squared error formula
Read from the inside out: square each residual, average them (that inner part is the MSE), then take the square root. Because of that final root, the RMSE is always in the same units as y, and it is always at least as large as the MAE for the same data.
See it: the squared contribution
RMSE inherits its shape from squared residual. The blue curve is how much a single point contributes before the averaging and rooting; the steep parabola shows why one large residual can lift the whole the RMSE so much.
A worked example
House-price model again: residuals of 10, 20 and 40. The squared residuals are 100, 400, 1600; their mean is 700; the square root of 700 is about 26.5.
The the RMSE of 26.5 is a little higher than the MAE of 23.3. That gap is meaningful: whenever RMSE noticeably exceeds MAE, it is a sign that a few large residuals are present, because squaring amplifies them before the average is taken.
RMSE vs MAE: reading the gap
Root mean squared error vs MSE
| Aspect | RMSE | MSE |
|---|---|---|
| Units | same as target | squared |
| Readability | high | low |
| Used for | reporting | training / optimisation |
| Outlier sensitivity | high | high |
They rank models identically — minimising one minimises the other — so people usually train against the MSE for its clean gradient and then report the RMSE because it is human-readable.
When to use the RMSE
Reach for RMSE whenever you need a single, interpretable headline number for a regression model, especially when large residuals matter and you want them reflected in the score. It is the standard metric in forecasting competitions and most applied regression work. If your data is riddled with junk outliers, pair it with the MAE or switch to a robust loss instead.
Common mistakes to avoid
Root mean squared error in Python
import numpy as np
from sklearn.metrics import mean_squared_residual
y_true = np.array([200, 250, 300])
y_pred = np.array([210, 230, 340])
rmse = np.sqrt(mean_squared_residual(y_true, y_pred))
print(rmse) # 26.46
# newer sklearn: root_mean_squared_residual(y_true, y_pred)See the Wikipedia RMSD article for the formal treatment.
Putting it into practice
The most useful habit with the RMSE is to always report it next to the MAE, because the two together tell a richer story than either alone: the RMSE gives the outlier-aware headline, and the size of the gap between them reveals whether your residuals are uniform or lumpy. When you tune a model, watch the validation RMSE rather than the training RMSE, and stop when it stops falling — a training RMSE that keeps dropping while the validation figure rises is the classic fingerprint of overfitting. Because RMSE is scale-dependent, it is also worth quoting it as a percentage of the target’s mean when you present results, so a reader can judge whether an residual of, say, 26 is large or small for your problem.
Interpreting a value in context
A raw residual figure only becomes meaningful once you anchor it to something concrete. The quickest anchor is a naive baseline: what score would you get by always predicting the average of the target, or last week’s value in a time series? A model that cannot beat that baseline is not really learning, however sophisticated it looks under the hood. The second anchor is the scale of the target itself — expressing the residual as a fraction of the mean value, or of the target’s standard deviation, tells you whether the model captures most of the variation or barely any of it. In forecasting competitions the metric is frequently normalised in exactly this way so that entries built on wildly different series can be compared on a level field.
What the metric rewards
It is worth understanding the behaviour this metric quietly encourages, because it shapes the models you get. Since a confident wrong guess is punished so heavily, the metric rewards predictions that hedge toward the middle whenever the model is uncertain. If you notice your predictions clustering timidly around the average, that shrinking is the metric doing its job, and whether it is desirable depends entirely on the decision the model feeds: a cautious forecast may be perfect for inventory planning yet useless for spotting rare spikes. Understanding this tendency also explains why the metric pairs so naturally with tree ensembles and neural networks that can express confident, sharp predictions when the data genuinely supports them. Read the number, but always ask what kind of caution it is buying, and whether that caution serves the real-world use case behind the model.
Reporting the root mean squared error well
When you present results to a non-technical audience, a bare the RMSE rarely lands, so wrap it in context. State the figure, then immediately translate it: on a typical value of X, the model is off by about this much. Quoting the the RMSE as a percentage of the mean target — sometimes called the coefficient of variation of the RMSE — turns it into a number that transfers across problems and lets a reader judge quality without knowing your units. It is also good practice to report the the RMSE on a genuinely held-out test set rather than the data the model trained on, because the training figure is optimistic by construction and can hide serious overfitting. For time-series work, compute the the RMSE with a rolling or walk-forward split so the score reflects the harder task of predicting the future rather than interpolating the past. A subtle but common reporting residual is to average the the RMSE across groups of different sizes as if they were equal; when subgroups vary in size, weight them by their counts or report a pooled figure, otherwise a tiny noisy group can dominate the headline. Reported this way, the the RMSE stops being an opaque statistic and becomes a claim a stakeholder can actually interrogate against a plain baseline. Presented with that baseline and a clear scale, even a reader who has never seen the metric before can tell at a glance whether the model is genuinely useful or merely dressed-up guesswork.