Model-dependent vs Robust predictors of risk tolerance: a dual-model variable importance approach

Empirical analyses of investor risk tolerance typically rely on a single modelling framework, leaving open the question of whether the identified determinants are genuine empirical regularities or artefacts of specification choices. This paper addresses this question by proposing a dual-model framework that combines an interpretable ordered probit model with an ordinal forest, a machine learning method designed for ordinal outcomes, and applies it to survey data from 1,549 investors in Finland, Germany, Italy, and Spain. Rather than selecting a single best-performing model, the framework uses the two approaches jointly to distinguish determinants that are robust to modelling assumptions from those that are model-dependent. To enable a direct comparison of variable relevance across models, we introduce likelihood-based and permutation-based measures of variable importance evaluated within a unified five-fold cross-validation procedure and scored using the Ranked Probability Score, a proper scoring rule for ordinal outcomes. The results reveal a clear robustness gradient. Age and country of residence emerge as the only high-robustness determinants, with stable importance rankings across all measures and negligible fold-to-fold variability. Education, occupation, and income form a second tier of consistently relevant predictors. By contrast, gender and children display notable divergence across models, indicating that their estimated effects are partially sensitive to functional-form assumptions. Financial literacy, marital status, and home ownership show consistently low and unstable importance across all measures. These findings have direct implications for portfolio suitability assessment: determinants with high robustness provide a reliable basis for investor classification, while model-sensitive variables introduce classification risk, the probability that the same investor receives different risk category assignments under alternative specifications.