Comparative Evaluation of Feature Extraction Techniques in Margin Call Cascade Detection: Balancing Accuracy and False Alarm Rates

Abstract

Margin call cascades represent critical systemic vulnerabilities within modern financial markets, potentially triggering widespread liquidity crises through procyclical feedback mechanisms. This research conducts a comprehensive comparative evaluation of feature extraction techniques for detecting margin call cascade risks, focusing on the fundamental tradeoff between detection accuracy and false alarm rates. Through rigorous experimental analysis utilizing ensemble learning approaches, including Principal Component Analysis (PCA), XGBoost-based feature selection, and hybrid extraction frameworks, this investigation examines the performance of these methods across multiple classification algorithms. Empirical results demonstrate that hybrid feature extraction with Gradient Boosting achieves an ROC-AUC of 0.921; at the model-comparison operating point (§4.2), the false positive rate is 8.1%. Threshold optimization in §4.3 yields FPR values ranging from 4.6% to 8.5%, depending on the criterion (F1/Youden/F2). The findings provide actionable guidance for regulatory authorities seeking to calibrate early warning mechanisms that strike a balance between timely risk detection and operational efficiency constraints.