Research

Abstract:

Abstract: Peer firm identification traditionally relies on proximity measure among firms, often based on static industry classification systems like GICS or text similarity from supply-side data like product descriptions in 10-K. Although recent studies have introduced dynamic methods to identify peers for a focal firm, they often suffer from issues related to data coverage, flexibility, and suboptimal performance. To address these limitations, we construct a firm-firm network based on co-mentions in posts from a large-scale social media dataset. Different from methods that use network linkages or typical structural measures (e.g., centralities), we employ a network representation learning technique that captures latent structures among firms, facilitating global structure learning beyond first-degree connections by projecting the network into a lower-dimensional space. Peer firms are identified via cosine similarity between learned embeddings of firms. To enhance the performance of peer firm identification, we develop an ensemble method that combines multiple sources of firm similarities. The results from the extensive evaluations demonstrate that our method outperforms existing methods in explaining cross-sectional variations in stock returns, financial ratios, and valuation multiples for base firms. Additionally, it offers several other desirable properties for peer firm identification. In summary, this study makes three key contributions: (1) leveraging the wisdom of crowds from social media to identify peer firms, enabling highly fine-grained analysis, (2) introducing a network representation learning technique and developing an ensemble method to combine multiple data sources for enhancing peer firm identification, and (3) extensively validating the proposed approach using a set of well-established benchmarking methods and showcasing its superiority through additional practical case studies.

Abstract: Assessing a firm’s credit rating is a significant and complex problem with wide implications for investors, markets and supply-chain partners of a firm. Typical approaches used in this context use financial, non-financial, and macroeconomic data as determinants for firm credit ratings. However, these are primarily tabular and fail to capture potentially relevant relational dynamics that exist in this ecosystem (firm-firm networks, analyst/board member relationships, etc.). This paper presents a novel heterogenous graph attention network (HGAT) with a multiplex corporate graph that integrates relational information from firm’s supply chain, director, and analyst networks, and then ensemble the learned embeddings with financial ratios for final credit rating predictions. The attention mechanism used facilitates interpretability and allows us to identify how different types of entities influence the final credit prediction of focal firms. Methodologically, we demonstrate that incorporating high-order topology structures at the motif level yields better results with explainable motif-based graphs. These contributions add to the growing literature in the information systems and finance areas on fintech innovations in markets, and have the potential of creating significant efficiencies given the widespread use of firm credit ratings in financial markets.

Abstract: Text-to-image (T2I) diffusion models have demonstrated impressive capabilities in generating visually appealing high-quality images. However, their performance highly depends on the quality of input prompts, making alignment between generated images and user intent a challenge. Existing approaches typically address this misalignment by leveraging LLMs for automatic prompt refinement or by fine-tuning diffusion models through RLHF. Although effective, these methods often require large-scale human-annotated training data or involve substantial computational costs. To overcome these limitations, we propose MAPO, a novel Multi-Agent Prompt Optimization framework designed to convert raw user prompts into optimized prompts tailored specifically for diffusion models, thus improving image synthesis quality aligned with user intentions. MAPO consists of four specialized agents — a refining agent, a T2I agent, an assessment agent, and a text-image alignment agent — that collaboratively and iteratively refine prompts, evaluate outputs, and optimize image generation quality. Empirical evaluations on three benchmark datasets show that MAPO significantly improves image aesthetics, text-image alignment, and human preference scores, outperforming several state-of-the-art training-based and training-free alignment benchmarks (an average improvement of 2.86%) in prompt-batch learning. In addition, MAPO achieves these advancements with minimal training data and limited computational resources, highlighting both its efficiency and effectiveness.

Abstract: While NLP techniques have advanced the automatic detection of political stance bias, little is known about effective visualization and communication of these biases. Drawing on cognitive psychology, this study investigates how local and counterfactual explanations influence users’ cognitive and metacognitive processes when evaluating biased news headlines. We fine-tune a bias detection model and conduct experiments to examine how different explanation designs based on SHapley Additive exPlanations (SHAP) affect user perception, reasoning, and decision quality. We hypothesize that interpretability designs can trigger different cognitive and metacognitive processes, such as confirmation bias and cognitive dissonance, which in turn shape user judgments. The research contributes to understanding how cognitive theories can enhance the design of AI systems, offering theoretical support for developing effective news bias mitigation tools. It provides practical insights for media platforms and policymakers seeking to enhance transparency, reduce cognitive burden, and promote bias awareness in digital online platforms.