Calibration of medium‐range temperature ensemble forecasts dependent on the time horizon: An assessment

Abstract The model conditional processor (MCP), a statistical post‐processing method introduced earlier for river flow forecasting, is evaluated for ensemble temperature predictions. Here, emphasis is put on time‐horizon dependence (MCP‐TD) by accounting explicitly for temporal correlations across forecast lead times. The processor models the dependence between ensemble forecasts and observations as a Gaussian copula, derives conditional univariate predictive distributions directly from data samples, and yields non‐parametric predictive distributions. The acknowledgment of the temporal correlation structure over a sequence of forecast steps makes MCP‐TD particularly appealing for calibrating forecasts of stongly autocorrelated variables like temperature. The performance of MCP and MCP‐TD is assessed against non‐homogeneous Gaussian regression (NGR). Both MCP variants produce forecasts with better reliability compared with NGR, while NGR yields sharper predictive distributions. Overall, the time‐horizon‐independent MCP and NGR achieve comparable performance. The proposed approach also attains the lowest continuous ranked probability score (CRPS) for daytime hours and shorter lead times, while its advantage diminishes for longer ones, where NGR outperforms it. For temporal dependence, the p‐variogram scores are comparable across nearly all methods, with COSMO‐LEPS being the only notable exception, indicating that the temporal correlation structure is preserved by the post‐processing methods. These encouraging results show a viable alternative to established post‐processing techniques in weather forecasting.