Jump Prediction in Systemically Important Financial Institutions' CDS Prices

Abstract

This research proposes a novel hybrid LSTM-KNN-GNN architecture for predicting price jumps in credit default swap (CDS) markets of systemically important financial institutions (SIFIs). The model integrates Long Short-Term Memory networks, K-Nearest Neighbors algorithms, and Graph Neural Networks through an innovative feature fusion mechanism, enabling simultaneous capture of temporal dependencies, local patterns, and institutional interconnectedness. The empirical analysis employs high-frequency CDS data from 25 global SIFIs spanning 2018-2020, encompassing over 750,000 observations. The hybrid architecture demonstrates superior performance in jump prediction accuracy, achieving improvements of 15.3% compared to traditional approaches. The model's effectiveness is particularly pronounced during periods of market stress, with ROC-AUC scores reaching 0.934 for one-day-ahead predictions. Network analysis reveals significant patterns in risk transmission, with highly connected institutions showing 23% higher susceptibility to price jumps. The research contributes to regulatory policy by providing empirical evidence on the effectiveness of network-based monitoring approaches. The analysis of cross-institutional spillover effects informs the development of macroprudential policy tools, with direct exposure channels accounting for 45% of observed spillover effects. The findings establish a robust framework for integrating machine learning techniques with traditional financial models while maintaining computational efficiency for real-time applications.