In decision theory, a scoring rule is a measure of performance of probabilistic predictions—made under uncertainty.
As an example of a probabilistic prediction, consider a sports magazine dealing with horse races that gives the winning chance of each horse for each race the day before. If we gather data regarding those predictions and compare it to the actual results, we have a measure – a scoring rule—of the magazine’s performance. This scoring is almost always nonlinear, and there are many different transformations which are widely used.
Proper scoring rules
A proper scoring rule is one that encourages the forecaster to be honest – that is, the expected payoff is maximized by accurately reporting personal beliefs about the predicted event, rather than by gaming the system. These rules include the Logarithmic scoring rule, Spherical scoring rule and Brier/Quadratic scoring rule.
The Brier score, for example, can be seen as a cost function. Essentially, it measures the mean squared difference between a set of predictions and the set of actual outcomes. Therefore, the lower the score, the better calibrated the prediction system is. It is a scoring rule appropriated for binary of multiple discrete categories, but it should be used with ordinal variables. Mathematically, it is an affine transformation of the simpler Quadratic scoring rule.
Logarithmic scoring rule
The logarithmic scoring rule assigns a negative payoff for all outcomes. The higher the score, the better calibrated the system is.
Score = log(abs(outcome—prediction))
where “outcome” is 1 or 0, and “prediction” is the probability on (0, 1) that the system assigned to the outcome that actually occurred.
References
Bickel, E.J. (2007). “Some Comparisons among Quadratic, Spherical, and Logarithmic Scoring Rules”. Decision Analysis, 4, (2), 49–65.
Tilmann Gneiting; Adrian E Raftery (March 2007). “Strictly Proper Scoring Rules, Prediction, and Estimation”. Journal of the American Statistical Association 102 (477): 359-378. (PDF)