Bayesian Recalibration of Upper-Level Wind Regime Indices

Upper-level zonal wind is a robust indicator for predicting episodes of consecutive dry days on subseasonal timescales in Norway. To construct a predictor that facilitates the predictive modeling of consecutive dry days, we define a regime-based index by projecting instantaneous upper-level flow anomalies onto three dominant circulation regimes identified using an agglomerative clustering algorithm. Next, we evaluate the skill of Integrated Forecast System (IFS) in predicting the regimes.

Results show that the raw probabilistic forecasts of the index are systematically over-dispered. To address this deficiency, we apply a Bayesian recalibration framework that combines a prior distribution derived from observed climatology with a likelihood function that represents the conditional dependence of observations on the ensemble mean. The resulting posterior distribution serves as the calibrated probabilistic forecast.

This modeling framework offers two key advantages: (1)Intrinsic Recalibration: when the forecast contains no usable predictive information and the likelihood converges toward the model climatology, the posterior automatically reverts to the observed climatology, ensuring well-calibrated probabilistic forecasts even in the absence of raw forecast skill (2) Sampling from complete samples: Bayesian method infers the posterior distribution efficiently from a shorter reforecast-reanalysis join sample (two times per week, 20 years), and a longer reliable reanalysis data (ERA5, ~40 years). This mitigates sampling limitations commonly encountered in non-Bayesian calibration methods, which rely solely on the joint forecast–observation sample and can therefore produce suboptimal distributional estimates, particularly when hindcast datasets are short.