Forecasting time to risk based on multi-party data: An explainable privacy-preserving decentralized survival analysis method

Abstract

Forecasting time-to-risk poses a challenge in the information processing and risk management within financial markets. While previous studies have focused on centralized survival analysis, how to forecast the time to risk using multi-party data in a decentralized and privacy-preserving setting with the requirements of explainability and mitigating information redundancy is still challenging. To this end, we propose an explainable, privacy-preserving, decentralized survival analysis method. Specifically, we transform time-to-risk forecasting into a multi-label learning problem by independently modeling for multiple time horizons. For each forecasting time horizon, we use Taylor expansion and homomorphic encryption to securely build a decentralized logistic regression model. Considering the information redundancy among multiple parties, we design and add decentralized regularizations to each model. We also propose a decentralized proximal gradient descent method to estimate the decentralized coefficients. Empirical evaluation shows that the proposed method yields competitive forecasting performance and explainable results as compared to benchmarked methods.