Empirical Validation of Machine Learning Based Financial Decision Support: Three Case Studies on Loan Risk Control, Corporate Credit Rating, and Stock Price Forecasting

Abstract

Over the past two decades, the rapid development of machine learning (ML) technologies has significantly transformed the methodologies and applications of financial research and practice, providing unprecedented power to extract novel insights from complex, multidimensional data. Unlike traditional statistical models, which usually depend on strict assumptions of linearity and normal distribution, ML technologies, ranging from classic logistic regression and decision trees to deep learning (DL) methods such as neural networks and transformers, have demonstrated advanced abilities to explore complex relationships, handle alternative data, and improve prediction accuracy across various financial domains. In addition to these advantages, the diverse applications of ML technologies not only address the core challenges in financial decision-making but also demonstrate astonishing capabilities for addressing the inherent heterogeneity, volatility, and information asymmetry of financial markets. Against this background, this dissertation aims to contribute to novel applications of ML and DL technologies in the following three financial cases: loan risk management, corporate credit ratings, and equity price prediction. Through rigorous comparative model analyses and empirical validations, this study provides valuable insights into both academic and practical applications of ML technologies for addressing real-world financial challenges.

The first case aims to solve the loan risk management problem of small and medium-sized enterprises (SMEs) using classic ML technology that combines explainable artificial intelligence and portfolio optimization methods. Financial credit ecosystems function as complex computational social systems in which loan defaults manifest as behavioral anomalies that deviate from normative financial patterns. Consequently, transparent anomaly detection is imperative for reliable decision-making. This study proposes an integrated end-to-end framework for explainable credit risk modeling and multiobjective portfolio optimization that is specifically tailored to SMEs. Initially, a logistic regression model is used estimate the probability of default (PD), and Shapley additive explanations is used to provide global and local explainability. This methodology facilitates a better understanding of critical risk drivers. Crucially, these explainable PD estimates are directly incorporated into a multiobjective portfolio optimization model. Using the evolutionary Nondominated Sorting Genetic Algorithm III (NSGA-III), this downstream step mathematically balances portfolio revenue, expected credit loss, and diversification. Experimental results validate that the proposed framework not only maintains robust predictive performance but also translates explainable risk insights into actionable, regulatory-compliant portfolio strategies. Ultimately, this study bridges the gap between explainable artificial intelligence and operations research. Specifically, it offers a transparent, datadriven paradigm for risk management in financial social systems by articulating how scenario analysis links accuracy, explainability, and portfolio optimization for the future integration of affective signals as alternative data.

The second case introduces a novel multimodal data fusion framework to examine the impact of extreme public sentiment on corporate credit ratings. Departing from traditional binary or ternary sentiment classifications, our approach leverages a fine-tuned bidirectional encoder representations from transformers (BERT) model to categorize 3,839,916 Twitter (now X) posts into five distinct sentiment groups, ranging from extremely negative to extremely positive. By integrating these refined sentiment signals with firm-specific financial data for the target S&P 500 companies, we construct a comprehensive multimodal dataset that enables a more granular investigation of the interplay between public opinion and credit changes. Employing a suite of econometric techniques, including two-way fixed-effects panel regressions, ordinary least squares, system generalized method of moments, and generalized linear models, this study demonstrates that extremely negative sentiment exerts a statistically significant detrimental effect on credit ratings. By contrast, the impact of extremely positive sentiment remains largely insignificant. Robustness checks, including sensitivity analyses, lag effect examinations, reverse causality checks, and nonlinear analyses, confirm that the extreme facet of negative public sentiment primarily drives the adverse influence on credit ratings. By fusing DL-based textual analysis with traditional financial metrics, our study not only refines the measurement of public sentiment but also provides robust evidence of its dynamic implications for corporate financial stability. This multimodal data fusion approach paves the way for future research to incorporate additional social media streams and advanced language models, thereby enhancing predictive accuracy and deepening insights into the financial ramifications of public sentiment.

To address the low prediction accuracy caused by the inherent high noise and nonstationary characteristics of stock price series, the third case proposes a novel stock price prediction framework that integrates adaptive signal decomposition with multiscale feature extraction. The framework first employs a cross-validated adaptive signal decomposition (CVASD) module, a variational mode decomposition method adaptively optimized by a porcupine optimization algorithm, to decompose the original stock price series into a set of intrinsic mode functions with distinct frequency characteristics, effectively separating noise and multifrequency signals. Subsequently, the decomposed components are fed into a prediction network based on the Informer (Beyond Efficient Transformer for Long Sequence Time-Series Forecasting) architecture. In the feature extraction phase, this study designs a multiscale dilated convolution module (MDCM) to replace the standard convolution of the Informer, enhancing the models ability to capture short-term fluctuations and long-term trends by using convolution kernels with different dilation rates in parallel. Finally, the predictions from each component are integrated to obtain the final predicted values. Experimental results on three representative industry datasets (Information Technology, Finance, and Consumer Staples) from the U.S. S&P 500 index show that compared with several advanced baseline models, the proposed CVASDMDCMInformer framework demonstrates significant advantages across multiple evaluation metrics, including mean absolute error, mean squared error, and root mean squared error. The ablation experiments further validate the effectiveness of the two core modules, CVASD and MDCM. This study indicates that the proposed framework can effectively handle complex financial time series and thus provides a new solution for stock price predictions.

Date of Award	27 Jul 2026
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Yiu Keung Raymond LAU (Supervisor)