IPO Awareness and Retail Investors' Subscription Preferences: An Explainable Ordinal Machine Learning Framework (Computer and decision making.)

Summary

• The study analyses 176 UAE retail investors and 1,936 respondent-IPO observations pertaining to 11 IPOs.p.1
• The majority of respondents had no prior IPO experience (78.41%) and were not very knowledgeable about the market (68.75%).p.8
• Salik had the highest IPO awareness rate (56.82%) while Fertiglobe had the lowest (21.02%).p.8
• AE-OCB achieved the lowest MAE (0.9029), outperforming AE-OXGB v4 (0.9520) by 5.15%.p.9
• Bootstrap analysis with 600 resamples gave ΔMAE of −0.0449 for XGBoost and −0.0191 for CatBoost, with intervals containing 0, so awareness adds no significant predictive value.p.10
• Temperature scaling reduced log loss by 9.29% but lowered macro-F1 and balanced accuracy.p.12

Key insights

1. AE-OCB achieved the lowest ordinal MAE (0.9029) among the tested models.p.9
2. Low macro-F1, balanced accuracy and quadratic weighted kappa indicate limited absolute predictive discrimination.p.9
3. Removing awareness predictors did not significantly change error, as both bootstrap intervals contained 0.p.10
4. Risk-neutral respondents had the lowest MAE (0.8466), while IPO-aware respondents had a higher MAE (1.0112) than IPO-unaware respondents (0.8603).p.10
5. In leave-one-IPO-out testing the lowest MAE (0.4261) was recorded for Burjeel and the highest (0.8239) for ADNOC Drilling.p.11
6. Risk attitude (0.0110) and focal awareness (0.0026) had small positive permutation importance, while other predictors were negative.p.11
7. Temperature scaling reduced log loss by 9.29% while macro-F1 and balanced accuracy decreased, showing a calibration-discrimination trade-off.p.12
8. Tecom achieved the highest categorical accuracy (41.48%) and ADNOC Drilling the lowest (24.43%).p.13