Skip to content
Calcrivo

Root Mean Squared Log Error (RMSLE) Calculator

Calculate Root Mean Squared Log Error (RMSLE), useful for regression targets spanning several orders of magnitude.

Inputs

Comma-separated actual (true) non-negative values.

Comma-separated predicted non-negative values, same order and count as actual.

RMSLE

0.112352

Number of Samples

4

Step by step

  1. Sum of squared log errors: Σ(ln(1+ŷ_i) − ln(1+y_i))²

    (ln(1+110) − ln(1+100))² + (ln(1+230) − ln(1+250))² + (ln(1+950) − ln(1+1000))² + (ln(1+60) − ln(1+50))²

    = 0.050492

  2. RMSLE: √[Σ(...)² / n]

    √(0.050492 ÷ 4)

    = 0.112352

How it works

Root Mean Squared Log Error (RMSLE) applies RMSE to the log-transformed values: RMSLE = √[(1/n) × Σ(ln(1+ŷ_i) − ln(1+y_i))²]. The +1 offset avoids taking the log of zero. Because it operates on logarithms, RMSLE treats relative differences symmetrically — an error where the prediction is half the actual value is penalized similarly to a prediction that's twice the actual value — making it well-suited for targets that span several orders of magnitude, like sales volume, population counts, or pricing data.

Formula

RMSLE = sqrt((1/n) * sum((ln(1 + y_hat_i) - ln(1 + y_i))^2))

n
Number of samples
y_i
Actual value for sample i
\hat{y}_i
Predicted value for sample i

Frequently Asked Questions

Why use RMSLE instead of RMSE?

RMSLE penalizes relative (percentage-like) errors rather than absolute ones, so it's less dominated by a few very large-magnitude targets — this makes it a better fit when the target variable spans several orders of magnitude, such as sales counts across products of vastly different popularity.

Why add 1 before taking the logarithm?

Adding 1 avoids taking the logarithm of zero (which is undefined/negative infinity) when actual or predicted values can legitimately be zero, a common situation in count data like sales or web traffic.

Does RMSLE penalize under-prediction and over-prediction equally?

Yes, roughly — because it operates in log-space, RMSLE treats a prediction that's proportionally too low similarly to one that's proportionally too high by the same ratio, unlike RMSE, which is symmetric in absolute rather than relative terms.

You might also need