Comparing XGBoost and ARCH-LM Models for IPO Valuation in the Iranian Capital Market

The main objective of this research is to conduct a rigorous comparative analysis of the performance of the eXtreme Gradient Boosting (XGBoost) algorithm against the traditional Generalized Autoregressive Conditional Heteroskedasticity (ARCH-LM) model in predicting initial public offering (IPO) valuation ratios in the Iranian capital market. This study focuses on three key dependent valuation variables: the market-to-book value ratio (m2b), the enterprise value-to-assets ratio (v2a), and the enterprise value-to-sales ratio (v2s). Utilizing a cross-sectional dataset of 163 Iranian companies (42 listed and 121 over-the-counter) during the 2013-2023 fiscal period, the research employs a robust out-of-sample validation approach. Performance is benchmarked using prediction accuracy, economic significance, and computational efficiency. Diebold-Mariano tests and bootstrapping were applied to ensure the statistical significance of predictive differences between the models. The comparative analysis revealed that the XGBoost model demonstrates statistically significant superiority over the ARCH-LM model across all predictive metrics. The out-of-sample (OOS) Coefficient of Determination (R2) showed an improvement ranging from 9.7% to 10.9% across the valuation ratios, with a Root Mean Square Error (RMSE) reduction between 32.6% and 35.5%. The OOS R2 values for XGBoost ranged from 0.715 to 0.768, with all predictive differences confirmed as significant (p