Recalibrate prediction intervals

Applies post-hoc recalibration to improve the coverage properties of prediction intervals from forecast. Two methods are available: isotonic regression (nonparametric, monotonicity- preserving) and Platt scaling (logistic, parametric).

Usage

recalibrate(forecast_obj, bt, method = c("isotonic", "platt"))

Arguments

forecast_obj

An lfq_forecast object to recalibrate.

bt

An lfq_backtest object providing the calibration data. The mapping from nominal to empirical coverage is learned from the backtest residuals.

method

Recalibration method: "isotonic" (default) for isotonic regression on the empirical coverage function, or "platt" for Platt scaling via logistic regression.

Value

An lfq_forecast object with recalibrated .lower and .upper bounds. The object retains all attributes of the original forecast and can be passed to autoplot.

Details

Isotonic regression: Learns the monotone mapping from nominal coverage levels to observed coverage using the backtest data, then inverts it to find the nominal level that achieves the desired empirical coverage. This is a nonparametric approach that requires no distributional assumptions.

Platt scaling: Fits a logistic regression of the form \(P(\text{covered}) = \text{logit}^{-1}(a \cdot z + b)\) where \(z\) is the standardised residual, and uses the fitted model to adjust prediction interval widths.

Both methods require a backtest object with sufficient data to estimate the calibration mapping reliably (at least 30 forecast-observation pairs recommended).

References

Platt JC (1999). Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large Margin Classifiers, 61–74.

See also

calibrate for calibration diagnostics.

Examples

# \donttest{
sim <- simulate_dynamics(n_lineages = 3,
  advantages = c("A" = 1.2, "B" = 0.8),
  n_timepoints = 20, seed = 1)
fit <- fit_model(sim, engine = "mlr")
fc  <- forecast(fit, horizon = 21)
bt  <- backtest(sim, engines = "mlr",
  horizons = c(7, 14), min_train = 42)
fc_recal <- recalibrate(fc, bt, method = "isotonic")
# }