Dependence-Aware Forecast Combination under Temporal Distribution Shift: European Insurer Solvency-Ratio Deterioration
Year: 2026
Type: preprint
Abstract: We evaluate probability forecast combination for one-year-ahead solvency-ratio deterioration using 9,239 labelled firm-years from European insurers. Model fitting, feature selection and calibration follow calendar order. Using three selected features in the primary evaluation, logistic regression and a linear support-vector machine achieve an out-of-time area under the receiver operating characteristic curve (AUC) of 0.727, versus 0.718 for Dempster fusion and 0.724 for the current solvency ratio alone. With all eight candidate features, Dempster fusion reaches AUC 0.735, exceeding the best single model by 0.008. Exploratory calibrated generalised least-squares pooling improves AUC by 0.01 over the validation-selected constituent; decision-focused pooling reaches AUC 0.733. Pooling results cover two test origins, with firm-clustered uncertainty conditional on those years. A counterexample and synthetic experiments illustrate how repeated evidence can produce overconfident probabilities. Ranking, calibration and operational review value require separate assessment. Independent temporal validation is needed to establish a general advantage for dependence correction.
Source: Research Square
Institutions: University of Rzeszów, Krakow University of Economics
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7226035140
