The Cost of Coverage: Conformalising Native Forecast Uncertainty for GB Wind Power

Modern wind forecasts publish uncertainty bands but actual coverage is rarely guaranteed, exposing grid operators to costly balancing actions. We conformalise the native bands of six models, from ARIMA to tabular foundation models, under real operational constraints, measuring what the wrapper adds and what it costs. The wrapper fixes marginal coverage at a width cost proportional to the native coverage deficit, with Winkler Scores holding flat or improving. Black-box forecasters gain a reliability floor without retraining, however, the guarantee is only marginal, and conditional coverage collapses in high winds when the weather forecast misclassifies the state. Crucially, these failures carry a reduced financial penalty because high-wind misses settle at the cheapest imbalance prices. We release the pipeline as an open benchmark.