DEEP LEARNING FOR FINANCIAL ANALYTICS: NETWORK REPRESENTATION LEARNING AND AI ALIGNMENT

Abstract

This dissertation develops deep learning approaches for Fintech applications, with a focus on learning and aligning representations from unstructured financial data to support high-stakes decision-making. It addresses key problems including peer firm identification, credit risk prediction, and the adaptation of large language models (LLMs) for financial contexts. By integrating large-scale unstructured data from social media, multiplex corporate networks, and earnings calls, the studies introduce novel approaches that capture firm relationships and generate financially meaningful insights with improved accuracy and interpretability.

The first chapter introduces a data-driven approach to identify peer firms using a firm-firm network based on co-mentions in social media posts. Unlike traditional methods that rely on static industry classifications, we utilize network representation learning to capture latent firm-firm structures, where peers are then identified through firm embeddings. The method outperforms existing techniques in explaining cross-sectional variations in monthly returns and financial metrics. We also conduct interpretability analyses to reveal the encoded financial ratios in the learned embeddings and the firm similarities as a pair-trading strategy for positive alpha.

The second chapter extends network representation learning to a heterogeneous context with multiple node and edge types. We develop a multiplex network approach that incorporate firms’ supply chain (SC), director, and analyst relations, augmenting traditional financial ratios with relational dynamics. The attention mechanism highlights the relative importance of entities (SC partners, SC customers, directors, or analysts) in driving predictions, while motif-based GNNs capture high-order network structures, providing interpretable explanations for their impact on credit ratings.

The third chapter focuses on LLM alignment to generate analyst-style reports that demonstrate finance-specific reasoning capabilities. We introduce domain-specific evaluation metrics capturing sentiment, language style, and content coverage, and show that baseline LLMs fail to reflect critical domain nuances. To address this, we propose a multi-objective alignment approach that incorporates linguistic signals, significantly improving report alignment to human language. The aligned models achieve predictive performance comparable to human analysts and provide incremental value for explaining and forecasting market reactions.