Log Loss
Definition
Logarithmic loss evaluates probabilistic classifications and strongly penalizes confidently wrong predictions.
For binary classification, Log Loss corresponds to the negative log-likelihood of the Bernoulli/logistic model; multiclass variants use cross-entropy over class probabilities.
Simply explained
How to penalize wrong answers especially strongly when the model was very confident.
Why it matters in practice
Relevant for classification, probabilistic models, and calibration.
Example from business
Two classifiers with the same accuracy differ in Log Loss because one is much more confident in its errors.
Limits and pitfalls
- Only consider accuracy; interpret probabilities uncalibrated; log numerical zero/one values unclipped.
Origin & history
Log Loss follows directly from maximum likelihood training of probabilistic classification models.
Use in mid-sized companies
- Possible application in SMEs: jointly evaluate probabilistic classifiers with accuracy and calibration metrics.
Frequently asked questions
Is lower Log Loss better?
Yes. Is Log Loss the same as Accuracy? No.
Sources
- log_loss · scikit-learn, log_loss
Sources checked on 10.10.2026.
Editorially reviewed on: