Multimodal Data Augmentation-enhanced Deep Learning Framework for Corporate Credit Risk Assessment

Description

On 23rd February 2022, the Financial Secretary of HKSAR has delivered his 2022-23 budgetspeech and stressed government’s aim to promote continuous innovation by encouragingresearch institutions to put forward Fintech solutions so as to remove development bottlenecksfaced by the Fintech industry in Hong Kong.1 Accordingly, the proposed GRF project isstrategically aligned with the HKSAR Government’s calls for technological innovation andFintech development in Hong Kong and the Greater Bay Area (GBA). More specifically, theproposed GRF project aims to develop a novel multimodal data augmentation-enhanceddeep learning framework to improve corporate credit risk assessment. With the increasing financial turbulence caused by the COVID-19 pandemic and globaleconomic recession, it has become increasingly more important to promptly assess the creditrisk of individual firms to avoid the nation-wide financial distress. Although there have beensome studies examining the topic of corporate credit risk assessment before, traditionalmethods mainly focus on quantitative financial factors alone, and qualitativeexplanations/justifications of these models’ assessments are absent. Recently, researchers havestarted to examine qualitative factors such as the high dimensional “sentimentalexpressions” of firms extracted from financial news articles, firms’ annual reports, and publicsocial media text to enhance corporate credit risk assessment. Accordingly, this GRF projectaims to exploit state-of-the-art artificial intelligence (AI) method, namely deep learningwhich is able to model complex nonlinear relationships between high dimensional inputfeatures (both quantitative and qualitative features) and output classes to bootstrap corporatecredit risk assessment. Nevertheless, one common problem of deep learning is the lack oflabeled training examples to train deep neural networks characterized by millions of modelparameters. Another problem is that the black box decision-making approach of deep neuralnetworks makes it difficult to explain these networks’ decisions/predictions to business users. Guided by the design science (DS) research methodology, the goal of the proposed GRFproject is to solve the aforementioned research problems. For instance, the proposed dataaugmentation-enhanced deep learning methods can alleviate the problem of lack of labeledtraining examples, and the multimodal data (e.g., textual sentimental expressions andsentimental video frames) can improve the “explainability” of deep learning-based corporatecredit risk assessment. So, this study contributes to develop Explainable AI as well. To ensureshort/long term business and societal impacts, we plan for technology transfer of ourresearch outcomes (e.g., a prototype system) through our industry collaborator that is listed onNasdaq. Because the proposed GRF project will develop a novel deep learning-based Fintechservice for the region, it contributes to develop Hong Kong as a financial technology hub in theworld, and thereby boosts Hong Kong’s economic growth in the long term.