Decompose Brier Score — decompose

Decomposes Brier score into reliability (calibration), resolution (discrimination), and uncertainty components.

Usage

decompose_brier(probs, labels)

Arguments

probs: Numeric vector of predicted probabilities
labels: Binary labels (0/1)

Value

A list containing: - brier: Total Brier score - reliability: Calibration component (lower is better) - resolution: Discrimination component (higher is better) - uncertainty: Inherent uncertainty (fixed for given class balance)

Details

Brier score decomposes as: $$Brier = Reliability - Resolution + Uncertainty$$

- **Reliability** (calibration): measures deviation from perfect calibration - **Resolution**: measures how much predictions differ across groups - **Uncertainty**: base rate of positive class, p(1-p)

A well-calibrated model has low reliability and high resolution.

Examples

if (FALSE) { # \dontrun{
probs <- c(0.1, 0.4, 0.6, 0.9)
labels <- c(0, 0, 1, 1)
decompose_brier(probs, labels)
} # }