TLDR: This research introduces two deep neural network models, DeNN and DSNN, designed to minimize Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) in peer-to-peer (P2P) loan portfolios. The models predict both the probability and timing of loan defaults to determine return distributions. Experiments on real-life Lending Club data show that both DeNN and DSNN significantly reduce portfolio VaRs compared to benchmarks. Notably, the simpler DeNN model often outperforms the more complex DSNN, providing a practical approach to P2P lending risk management.
Peer-to-peer (P2P) lending has emerged as a significant force in the financial industry, offering an alternative to traditional banking by connecting borrowers directly with lenders. Platforms like Lending Club have facilitated billions in loans, reducing transaction costs and potentially offering higher returns for lenders and lower costs for borrowers. However, with these opportunities comes a critical challenge: managing the risk of loan defaults. For individual investors, scrutinizing every loan is impractical, making portfolio diversification a common strategy.
This research paper, titled Minimizing the Value-at-Risk of Loan Portfolio via Deep Neural Networks, addresses this very issue. Authors Di Wang and Ye Du from Southwestern University of Finance and Economics propose innovative deep neural network models to help P2P lenders minimize the risk of their loan portfolios. The core idea is to predict not just whether a loan will default, but also when it might default, which is crucial for understanding the full distribution of potential returns.
Understanding Risk in Loan Portfolios
In finance, Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) are standard measures used to quantify the potential loss of an investment. VaR, for instance, tells an investor the maximum loss they can expect over a certain period with a given confidence level. To calculate VaR for a loan portfolio, the return distributions of its individual loans must be known. This paper introduces two deep neural network-based methods to predict these return distributions for P2P loans.
Introducing DeNN and DSNN Models
The first model, **Default Neural Networks (DeNN)**, is a low-degree-of-freedom model. It treats a loan’s return as a binary outcome: either the return under non-default or the return under default. DeNN employs two separate neural networks: one (DR-NN) to predict the default rate and another (DL-NN) to predict the default lifetime (how many months until default). By combining these predictions, DeNN estimates the loan’s return distribution.
The second model, **Deep Survival Neural Network (DSNN)**, is a high-degree-of-freedom model. Unlike DeNN, DSNN aims to predict the *distribution* of the default lifetime, meaning it estimates the probability of default at each installment period. This provides a more granular view of potential default times. DSNN is a novel two-branch deep neural network with a specially designed loss function and structure, incorporating survival analysis principles. One branch, the Survival Neural Network (SNN), estimates the hazard and survival functions, while another Deep Neural Network (DNN) acts as an ‘expert’ to supervise the SNN’s default rate prediction, enhancing its accuracy.
Minimizing Risk Through Simulation and Optimization
Once the return distributions for individual loans are predicted by either DeNN or DSNN, the paper outlines a Monte-Carlo simulation-based algorithm to minimize the VaR and CVaR of a portfolio of loans. This involves simulating thousands of possible return scenarios for a portfolio and then using an optimization algorithm (Sequential Least Squares Programming) to find the optimal allocation of capital (weights) among the loans to minimize the portfolio’s risk.
Also Read:
- Machine Learning Unlocks Enhanced Bank Bankruptcy Prediction
- DGTEN: A New AI Model for Reliable Trust Assessment in Dynamic Online Systems
Experimental Results and Key Findings
The models were tested on a real-life dataset of 244,720 loans from Lending Club. The experiments compared DeNN and DSNN against several benchmarks, including logistic regression models (LRM), simple survival neural networks (SNN), and basic diversification strategies like equal or random capital allocation. The results were compelling:
- Both DeNN and DSNN significantly outperformed the Equal weights and Random weights baselines in minimizing VaR and CVaR across different confidence levels.
- Compared to their respective benchmarks, DeNN and DSNN also showed considerable improvement. For example, DeNN reduced the VaR at 95% by 17.45% compared to LRM when optimizing CVaR at 95%. DSNN reduced VaR at 95% by 6.26% compared to a simple SNN when optimizing VaR at 95%.
- Interestingly, the DeNN model, despite its simpler design and lower degree of freedom, consistently outperformed the more sophisticated DSNN model in most scenarios. This suggests that sometimes a more straightforward approach can yield better practical results in complex financial modeling.
This research marks a significant step in P2P lending risk management, being among the first to use deep neural networks to predict the return distributions of P2P loans and to tackle the problem of loan portfolio VaR optimization. The findings offer valuable insights for investors and platforms looking to enhance risk management strategies in the booming P2P lending market.
