Deep learning for credit risk management under market complexity and illiquidity

Korangi, Kamesh (2025) Deep learning for credit risk management under market complexity and illiquidity. University of Southampton, Doctoral Thesis, 162pp.

Record type: Thesis (Doctoral)

Abstract

This thesis investigates three problems relating to the credit or financial risk
management of Small and Medium-sized Enterprises (SMEs) and listed mid-cap
firms. Both types of firms face specific challenges affecting their access to credit. Mid-caps firms have to deal with various market complexities and are often crowded out by large-cap firms, whilst SMEs have only indirect market exposure and collectively are a considerable risk to lenders. By utilising alternative data and recent advances in deep learning, the three papers forming this thesis develop and empirically test a series of novel prediction methods that can contribute to decreasing the cost of capital for these firms and enhance risk management practices for lenders.

First, in Chapter 1, an introduction is provided outlining the contextual setting of the thesis, the research aims and the intended contributions of the three papers.

The first paper (Chapter 2) is a study on default prediction for mid-cap firms, which introduces the challenges they face and modelling issues for predicting the default term structure. Different deep learning models are introduced and a novel multimodal architecture is proposed to make effective use of fundamental, market and pricing data, along with a framework to interpret the model predictions. The results show that deep learning models are powerful predictors and confirm some results from the literature.

The second paper (Chapter 3) studies large-scale time-varying portfolio optimisation for the same class of mid-cap firms. It shows how to filter complex networks at a large scale by combining existing techniques in a novel way. These networks are then used as inputs to a deep learning architecture that employs graph neural network models and a series of further layers for portfolio selection. The results confirm the effectiveness of using network information when devising portfolios to maximise return per risk, showing robust performance of the graph neural networks over long periods. Unlike earlier studies, this study shows that investing in peripheral firms in the networks might create additional risks. To our knowledge, this is the first study that includes firms that defaulted over the data period and explicitly considers
changes in the universe of investable firms over time.

The third paper (Chapter 4) studies credit lines of small and medium-scale enterprises to predict their default probability using behavioural and network data. Building on the previous two papers, we use a multimodal model with graph neural networks and deep learning to advance behavioural credit default prediction models. We use explicit networks from transactions, ownership and supply chain relationships over a large set of such firms and together with behavioural data that we derive from the revolving credit lines usage, we find the behavioural data highly predictive of default whilst the need for more complex models arises when using the network data.

Finally, Chapter 5 concludes with methodological contributions and the scope of
application of these studies, individually and collectively. It also puts forward ideas for future studies that could extend the application of deep learning models to other credit risk modelling problems.