ABSTRACT This study examines the determinants of the performance of initial public offerings (IPOs) conducted by 237 firms across 38 sectors on the Istanbul Stock Exchange between 2013 and 2025, using a hybrid analytical framework. To analyze the impact of sector characteristics, firm‐level financial indicators, supply structure, and market timing on the returns of USD‐denominated IPOs, this study combines heteroskedasticity‐resistant ordinary least squares (OLS) regression with tree‐based ensemble machine learning algorithms (random forest, XGBoost) and benchmark models (support vector regression, artificial neural network). The distributional characteristics of returns from IPOs highlight the use of nonlinear modelling techniques characterized by high positive skewness (5.12) and excess kurtosis (45.24). Our findings reveal significant sectoral heterogeneity. According to the findings, the Financial Services and Fisheries sectors showed statistically significant superior performance, while the Metal Mining, Insurance, and Construction sectors underperformed. Transportation and Storage exhibited the highest point estimate but did not reach conventional significance levels. The XGBoost machine learning model, with an R 2 of 0.869 and a mean absolute error (MAE) of 17.87, achieved superior single‐split prediction accuracy compared to the OLS method, which yielded an R 2 of 0.220. However, cross‐validation analyses indicated that all models exhibited limited out‐of‐sample generalizability (best CV R 2 = 0.04). The Diebold–Mariano test, conducted to compare prediction performance, confirmed that the artificial neural network (ANN) model provided a statistically significant improvement over the OLS method (DM = 2.121, p = 0.034). The SHAP value analysis used in the models reveals how each feature affects the final predictions, the relative importance of each feature, and the model's dependence on feature interactions. It identifies the IPO price (USD), the USD/TRY exchange rate, and the size of the supply as the most effective estimators with economically significant nonlinear interactions. The study aims to demonstrate the comparative advantages and limitations of ensemble learning under abnormal financial data conditions by presenting detailed sectoral evidence across 38 sectors. In this way, it aims to contribute to the IPO literature in emerging market economies by offering viable policy recommendations for investors, issuers, and regulators in Turkish capital markets.
Doruk Ayberkin (Thu,) studied this question.