CRPSScore<T>
Models & Types · AiDotNet.Scoring
Continuous Ranked Probability Score (CRPS) scoring rule.
Inheritance: ScoringRuleBase<T> → CRPSScore<T>
For Beginners
CRPS rewards predictions that are both accurate (close to the true value) and confident (narrow distributions). Unlike log score, CRPS:
- Has the same units as the predicted variable (like MAE)
- Is robust to outliers
- Considers the full shape of the distribution, not just probability at one point
For point forecasts, CRPS reduces to mean absolute error.
How It Works
CRPS measures the integral of the squared difference between the predicted cumulative distribution function (CDF) and the empirical CDF of the observation. It generalizes the mean absolute error to probabilistic predictions.
CRPS = ∫ (F(x) - 1(x ≥ y))² dx where F is the predicted CDF and y is the observation.
Constructors
| Constructor | Summary |
|---|---|
CRPSScore(int numIntegrationPoints) |
Initializes a new CRPS scoring rule. |
Properties
| Property | Summary |
|---|---|
bool IsMinimized { get; } |
|
string Name { get; } |
Methods
| Method | Summary |
|---|---|
T Score(IParametricDistribution<T> distribution, T observation) |
|
Vector<T> ScoreGradient(IParametricDistribution<T> distribution, T observation) |