Deep Learning for credit risk management under market complexity and illiquidity

Traditional models in credit risk have catered to individual consumers with statistical models and to corporates with structural or proprietary models like Moody's KMV. SME's (Small and Medium scale enterprises) have not been served well due to these divisions which restrict access to markets and availability of credit. Lack of structured data, ownership issues, legacy banking systems and limitations on the models themselves also contribute to this problem [1]. A fresh approach is needed to solve this problem, adapting models from different spheres and integrating the learnings in risk measurement techniques.
In a variety of application areas, a growing amount of high-dimensional data, increased computation power and new learning algorithms (deep learning) are allowing data scientists to solve complex problems. They are being applied in image and speech recognition, reconstructing brain circuits, new drug molecules discovery, natural language understanding and many other fields [2]. In this research, we focus on the recent advances in sequence models which were designed for natural language problems like language translation, sentiment analysis and vision problems, but have a natural application in time series data such as that used for market analysis and credit risk modelling. Temporal convolutional networks (TCN) outperformed a number of existing recurrent networks on standard tasks showing longer term memory [3]. Self-Attention models, which were designed for language translation tasks, are producing state of the art results and generalizes well [4]. Our proposal will take these state-of-the-art models as a base and develop better systems to understand complex risk structures.
The first research question aims at understanding the complexity of financial markets. Due to inter-dependencies, liquidity and behavioural challenges they are difficult to predict. Defining markets into transition regimes by categorizing on return, volatility and duration would improve systematic decision making, serving as inputs for other investment needs. We have partnered with Risk tech company, CheckRisk for this first study. They provide risk services to over $70bn of assets globally based on such market regimes. Using a combination of time series financial data from macroeconomic series, sector data and fundamentals creates a high-dimensional dataset that will serve as inputs to identify transition probabilities of different states of the market (e.g. scoring of risk from very positive to very negative). Improvements on these regimes will see an immediate impact in pricing market risk and aids in improving investment decisions in UK equity markets.
The second research question aims at understanding the interaction between credit and market risk [5]. This will allow us to better model credit risk of illiquid companies like small cap companies (Russell Index in US, FTSE 250 or AIM in UK) as well as further refine the model architectures for publicly traded SMEs. We will do so by incorporating market-related signals into their credit risk models, thus combining the types of models used in consumer lending with those in corporate credit risk modelling. To do so we will be needing the capability to model complex interactions inherent to deep learning models. By applying the sequence models developed in the first project component on market risk and by bridging the limitations of existing models, we will create a new approach towards measuring credit risk for SME's.
As a final research question, the regulatory impact of these models needs to be understood. As deep learning models are not particularly suitable for interpretability, confidence intervals on the network predictions are important. We propose to study Deep Bayesian networks [6] to arrive at financial confidence intervals, further breaching the gap between these advanced quantitative models and modern financial practice.