Journal of Risk Model Validation

Steve Satchell
Trinity College, University of Cambridge

Fall (autumn) is the time of year when research conferences flourish, and I have recently come back from one on private equity and credit. The overwhelming takeaway was how problematic data can be. Most opportunely, this subject is addressed in the first paper in this issue of The Journal of Risk Model Validation and, to a lesser extent, in the second. The issue’s third paper is concerned primarily with litigation risk: a very real risk for asset managers, and one that is rarely investigated. As an aside, managers tend to cultivate the image that all their clients are happy all the time, and most cases are consequently settled privately before going to court, creating a problem with respect to data collection.

The issue’s first paper, “A model combining Optuna and the light gradientboosting machine algorithm for credit default forecasting”, is by Xinyong Lu, Yuchong Li, Haoyan Wei, Jiaxin Wang, Xuewei Liu and Jiahui Wei. They note that the popularity of credit loans has increased significantly in recent years, which has caused credit default rates to steadily rise. Banks and other financial institutions have experienced significant economic losses as a result. However, the complexity, length, high dimensionality and imbalanced classes of real-world bank loan data make its analysis extremely difficult. In particular, these complications make it challenging to use standard credit default prediction models without some form of modification. This study proposes a credit default prediction model called Opt LightGBM to address this.

Our second paper has the somewhat lengthy title “Research on the multifractal volatility of Chinese banks based on the synthetic minority oversampling technique, edited nearest neighbors and long short-term memory”. Its authors, Jingyuan Huang, Yunhan Qu and Cheng Li, explore problems raised in risk validation exercises on China’s bank equity market when the data are deficient. They advocate a process that improves a long short-term memory (LSTM) neural network by combining it with the SMOTEENN method (the paper title gives the derivation of this acronym). They claim that fivefold cross-validation shows that, compared with the support vector machine and back propagation neural network models, their SMOTEENN-LSTM model has superior stability in accuracy indicators such as the area under the receiver operating characteristic curve and the geometric mean, indicating its strong robustness and reliability. This provides some evidence that it can deal with the problems encountered by traditional models when processing imbalanced data sets. This paper may well provide a useful procedure for financial regulators and financial institutions to better monitor and manage the risks of bank equity markets.

“Litigation risk assessment: a novel quantitative recency–frequency–monetary model” by Guodong Shi, Jianjie Huang, Jiahao Hou and Zeliang Zhang, the final paper in this issue, uses the recency–frequency–monetary (RFM) and Kealhofer– McQuown–Vasicek (KMV) models to assess companies’ litigation risk and credit risk, respectively. The authors employ the panel vector autoregression model to study the interrelationships between these risks. They start by proposing an RFM-based litigation risk assessment model. To validate the risk model and verify its robustness, they adopt the KMV model for accuracy testing and perform robustness checks using data sets for different periods. Their research findings indicate a correlation between the litigation risk calculated by the RFM model and the credit risk calculated by the classical KMV model. Specifically, companies with higher litigation risk tend to exhibit higher credit risk. Moreover, there is a bidirectional causal relationship between a company’s litigation risk and its credit risk, with each Granger-causing the other. This suggests that when evaluating a company’s overall risk, it may be sensible to consider both these risks in addition to other sources of risk.