Bridging Predictive Power and Interpretability: A Hybrid Deep Learning Framework for Real Estate Performance Under Systemic Uncertainty

Abstract This study develops an interpretable hybrid deep learning framework to forecast and explain the financial performance of real estate firms in emerging markets facing macroeconomic uncertainty. Using panel data from 28 listed Vietnamese firms during 2013–2025, a Macroeconomic Uncertainty Index (MUI) is constructed through principal component analysis to integrate five global risk indicators: economic policy, geopolitical, ESG-related, global uncertainty, and world sentiment indices. The proposed XGBoost–CNN model, augmented with SHAP explainability, achieves superior predictive accuracy, reducing RMSE by 54 percent and MAE by 22 percent compared with conventional deep learning models. SHAP analysis, applied to both the composite and individual uncertainty components, enhances interpretability and isolates the marginal influence of each risk factor. Results indicate that earnings per share and firm size jointly explain most of the variation in return on equity and assets, while economic policy uncertainty exerts the strongest macroeconomic impact. The MUI demonstrates a nonlinear moderating role, showing that, under heightened systemic uncertainty, traditional firm–performance relationships may reverse. The framework offers a transparent, data-driven tool for financial forecasting and macroprudential stress testing in emerging real estate markets.