IPO Awareness and Retail Investors' Subscription Preferences: An Explainable Ordinal Machine Learning Framework

Initial public offering (IPO) subscription preferences of retail investors must be predicted using computational methods that take investor heterogeneity, ordinal outcomes, and an informational characteristic into account. This study constructed an explainable ordinal machine-learning model to compare the incremental predictive power of IPO awareness and to test the reliability of financial decision-support. A total of 176 UAE retail investors and 1,936 respondent-IPO observations pertaining to 11 IPOs were analysed. Respondent-grouped nested cross-validation was used for the evaluation of Awareness-Enhanced Ordinal XGBoost (AE-OXGB v4) and Awareness-Enhanced Ordinal CatBoost (AE-OCB). Predictive accuracy, interpretability, generalizability and probability calibration were assessed by awareness ablation, subgroup analysis, leave-one-IPO-out validation, permutation feature importance and temperature scaling. AE-OCB had lower ordinal mean absolute error (MAE = 0.9029) compared to AE-OXGB v4 (0.9520) but overall discrimination of the classifications was low. When we only considered features without awareness, MAE for CatBoost was 0.8838 and for XGBoost was 0.9070, which shows that there was not enough incremental predictive power. Risk attitude and focal IPO awareness were two small, unstable positive permutation importance. The transfer performance was found to be heterogeneous from the results of Cross-IPO evaluation. Logarithmic loss decreased from 1.5508 to 1.4068 and the probability oriented performance improved, with the macro-F1 decreasing, as a result of the temperature scaling. The results suggest that informational relevance does not necessarily go hand-in-hand with improved prediction. The proposed framework integrates the potential of the ordinal modelling, the explainability and reliability assessment and provides methodological insights for financial decision support. But the amount of predictive discrimination (and sample size) will need to be validated independently for operational deployment.