Credit monitoring and proactive risk management with deep learning models

Credit risk monitoring is a long-overdue endeavor in the field of credit management, and this research sheds new light on how to proactively assess risk using deep learning methods on irregular time series. The greatest challenge is data irregularity, in which the timestamp of an event is often missing and its arrival can be delayed for a long time, which can result in silence intervals. Unlike methods developed in the financial domain, the shares of data discontinuity are studied in the field of event sequences to design a new simulation method of irregular time series data. A new architecture is also designed to address the challenge of irregular time series, incorporating specific embeddings, graph attention networks, and several attention modules based on multi-domain event information. This research ends with discussions on applications to other fields, the interpretability of the model's decisions, and risk monitoring from a financial-metric-centered perspective.

The proper prediction of credit risk is crucial to individual consumers, firms, and industries as a whole, especially with the development of peer-to-peer lending platforms and other new types of credit consumption. Overall, P2P is necessary and promising in China, which is supported by both individuals' credit scores and social networks. Additionally, since P2P companies lack supervised information, there's potential to utilize unstructured information in the financial arena or account social networks learned from overseers. P2P companies need new analytical and managerial capabilities regarding honoring anti-money laundering regulations, law enforcement sharing standards, and investigating high-risk customers with comprehensive territory analysis. Finally, consumers need to simplify their credit management processes, expectations of loan owing, and repayment terms.Credit risk prediction involves predicting the probability of default for a certain applicant on an ad-hoc loan. Either a deterministic number of seconds or an ordered series of seconds can be treated as a time series. Time series data from real-world FinTech companies can either be transformed to a non-sequential format after data preprocessing or work with the block box of models without clear causal relationships. During the past decade, multi-stage deep learning models with different mechanisms, illustrations and neural architectures including recurrent neural networks, long short-term memory, attention network and transformers have been devised to deal with the sequential and temporal nature of the data. Nonetheless, the overwhelming majority of existing works rely on widely adopted time series databases or simulated sequential datasets. Our work differed fundamentally from these existing works in financial credit risk prediction with sequential features in three aspects: unique modeling of features as linguistic text; multi-head design of dependency models.