<p>Corporate default prediction is crucial in risk management, yet traditional static models based on fixed windows hard to fully capture evolving economic dynamics. To address this gap, we propose a novel multi-window weighted random forest model explicitly integrating temporal proximity and historical predictive accuracy (measured by G-mean). We further implement a dual-directional feature selection method, systematically combining forward and backward selection processes to identify optimal feature sets. Employing the sample of Chinese listed companies from 2010 to 2020, our random forest-based default prediction model significantly outperforms traditional statistical methods (e.g., logistic regression), classical machine learning approaches (e.g., decision trees, K-nearest neighbor), and advanced deep learning models (e.g., artificial neural networks and deep neural networks) across critical predictive metrics including AUC, G-mean, and Recall. This paper further reveals that key features selection effectiveness of corporate default prediction varying across time windows. Our proposed method provides policymakers with effective tools to measure and capture default risk.</p>