AI Glossary / AI and machine learning / Intermediate
Log Loss
Definition
Log Loss, also called cross-entropy loss, is the standard error measure for probabilistic classifiers: it punishes wrong confident predictions severely, scoring how honestly the model's probabilities match the outcomes.
What it means
A classifier that says 90 per cent and is wrong should be marked down harder than one that says 55 per cent and is wrong. Log Loss formalizes exactly that: the penalty grows without bound as confidence in the wrong answer rises.
The score is built from logarithms of the predicted probability assigned to the true outcome: high probability on the truth, small loss; high probability on the wrong class, enormous loss. Averages over cases give the model's score.
The shape enforces honesty: hedged probabilities near the middle are penalized modestly when wrong, while certainty that fails is punished brutally. Models trained under Log Loss learn calibrated probabilities, not just labels.
It is both a training objective and an evaluation metric: most probabilistic classifiers are fitted by minimizing it, and model comparisons use it when the quality of the probabilities themselves matters, as in pricing, ranking and risk.
For a business, Log Loss answers a harder question than accuracy: not did the model pick the right side, but did it know how sure to be. Wherever decisions are sized by confidence, stakes, prices, thresholds, that harder question is the one that matters.
How it works
Have the model output a probability for the true class of each case, between zero and one.
Take the negative logarithm of that probability: near one means tiny loss, near zero means huge loss.
Average across cases for the model's score, with lower being better and zero being perfection.
Compare models on Log Loss when probability quality matters, and pair it with a calibration check for the full picture.
Watch for overconfident models: high accuracy with poor Log Loss signals certainty that the outcomes do not support.
Three practical examples
- A fraud model tuned to minimize Log Loss, so its 2 per cent scores really mean two in a hundred, usable directly in expected-loss arithmetic.
- A model bake-off where the accuracy winner loses on Log Loss, exposed as confidently wrong too often to trust with sized decisions.
- A training dashboard where Log Loss on validation data flattens while accuracy still climbs: the model has stopped learning honesty.
Think of it this way
Think of a weather forecaster paid by the accuracy of their percentages: rain said at 80 per cent on a dry day costs far more than the same miss said at 55 per cent. Log Loss is the bookkeeper of that scheme, settling up forecaster by forecaster, day by day.
The limit: the bookkeeper only scores what was said, not what should have been said. A model can game good averages while being wrong in a region nobody sampled, so the score still depends on honest, representative data.
Worked example
pricing with honest percentages
Suppose an insurer prices policies from a model's risk probabilities. The figures are illustrative.
Two candidate models tie on accuracy, but their Log Loss differs sharply: one says 90 per cent and is right 90 per cent of the time, the other says 90 per cent and is right 75 per cent.
Pricing from the second model's probabilities systematically underprices risk, because the confidence is borrowed, not earned.
The first model, the Log Loss winner, feeds pricing directly, and the comparison settles what accuracy could not.
Illustrative case study
A fictional betting syndicate trained a model for accuracy and staked proportionally to its confidence, discovering over a season that the confidence was systematically inflated.
The losses were mathematical rather than unlucky: staking by probabilities that overstate themselves loses by construction, and Log Loss would have announced the inflation before the first bet.
The invented lesson: when money is sized by confidence, measure the honesty of the confidence. Accuracy grades the picks; Log Loss grades the staking plan.
Common misconceptions and mistakes
- Accuracy is enough for probabilities. Accuracy ignores the size of the certainty; Log Loss exists precisely because 51 per cent and 99 per cent are different claims.
- Lower Log Loss always means a better business outcome. It means more honest probabilities; the business outcome still depends on how the probabilities are used.
- It is only for binary problems. The same scoring extends to many classes and to any setting where predicted probabilities are compared with outcomes.
Business use
Use Log Loss whenever probabilities drive sized decisions: pricing, limits, triage levels, ranking.
Pair it with a calibration plot in reviews: the score and the picture together show whether confidence is earned.
Treat an accuracy winner with poor Log Loss as a warning sign, not a winner, for any confidence-sized use.
People also ask
What does Log Loss actually measure?
The average negative log of the probability the model gave the true outcome, so confident wrongness is penalized far more than hedged wrongness.
Is cross-entropy the same thing?
Yes: the two names refer to the same loss, with cross-entropy common in engineering and Log Loss in evaluation.
Why does minimizing it calibrate probabilities?
Because the only way to score well on average is to make stated probabilities match true frequencies, which is calibration.
Check your understanding
Why does Log Loss punish confident wrong predictions so severely?
Answer: Because the penalty grows without bound as confidence in a wrong answer rises, so the only low-scoring strategy is to make stated probabilities match how often the model is actually right.
Related terms
Learn next
AI for Business Owners covers how to tell whether a model's confidence is earned before sizing decisions by it: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]
More in AI and machine learning
- Linear separability - Linear separability is the property that two classes of points can be separated perfectly by a linear decision boundary in the chosen feature space.
- Local case-control sampling - Local case-control sampling is a subsampling method for fitting logistic regression to large datasets.
- Local tangent space alignment - Local tangent space alignment, or LTSA, is a nonlinear dimensionality-reduction method that estimates local tangent spaces around data points and aligns...
- Logic learning machine - Logic Learning Machine is a rule-learning approach associated with Marco Muselli's switching neural network work and implemented in the Rulex software.
- Loss curve - A loss curve is the plot of a model's error score over the course of training: the falling, or sometimes worryingly not-falling, line that shows whether and...
- Loss surface - A loss surface is the landscape of a model's error over all possible parameter settings: the hills and valleys training must descend, whose shape decides...
Sources
[S2] Google, Machine Learning Glossary; [S14] KNIME, Data Science Glossary; [S19] scikit-learn documentation, Metrics and scoring. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Log-linear models Next: Log-normal distribution