Forecasting European insurance technical performance with global machine learning: rolling-origin evidence from a short panel

We forecast the technical performance of European insurance markets with a global machine learning approach. The data are a quarterly panel of 30 European countries on a common 2017Q1 to 2025Q4 grid. We target nine non-redundant technical indicators, including the claims ratio, the expense ratio, and the underlying growth flows. A single model is trained across all countries at each forecast origin, under a strict real-time rule. We evaluate with rolling-origin tests at horizons of one, two, and four quarters, using 19 to 23 forecast origins per target and horizon. Global machine learning beats a seasonal naive benchmark on most targets, with relative mean absolute errors from 0.64 to about 1.23. The gains are largest for the growth flows and weak or absent for the expense ratios, which we report plainly. We test equal predictive accuracy with origin-level Wilcoxon signed-rank tests, a benchmark we choose because the effective time-series sample is short, and we control the false discovery rate across the non-collinear target family. After both corrections, 65 of 243 comparisons remain significant. Global models also beat per-country autoregression and exponential smoothing. Against local learners the advantage is large for the elastic net, present for the extremely randomised trees, and mixed for the random forest. We add a single illustrative early-warning exercise. Catastrophe group, subgroup, and severity features do not improve detection of combined-ratio stress over a technical-only baseline. The test is powered against an area-under-the-curve gain of about 0.03 to 0.06, so it rules out a large catastrophe contribution.