Belinda Chen, University of Illinois-Urbana-Champaign
Abstract: Credit Default Swap (CDS) spreads exhibit network effects due to firms’ default interdependence. This paper employs Graph Neural Networks (GNNs) to predict CDS spreads by modeling firms as nodes and the measures of idiosyncratic volatility spillovers as directed edges. GNNs capture inter-firm network dynamics, improving prediction accuracy by over 50% compared to traditional models without edge features. We enhance the GNN with node-and edge-attention layers, identifying key nodes (e.g., manufacturing and intermediary firms) and edges (e.g., connections between intermediary, retail trade, or information firms and other firms) as critical to CDS spread prediction.
Discussant: Guillaume Roussellet, McGill University
Abstract: A key question in automating governance is whether machines can recover the corporate ob- jective. We develop a corporate recovery theorem that establishes when machines can do this. Training a machine on a large dataset of firms’ investment and financial decisions, we find that managers systematically underestimate investment costs, leading to over-investment and under-exploration. This bias persists even when accounting for intangibles, managerial compensation, and ESG scores. While social and governance concerns influence corporate ob- jectives beyond materiality, environmental concerns do not. Last, we observe that managerial alignment with shareholder value is imperfect, but it has improved over time.
Discussant: Ali Kakhbod, University of California-Berkeley
Abstract: The core technology powering modern Large Language Models (LLMs) estimates the distribution of probable answers conditional on the prompt. Using a financial news and returns dataset, we find that these conditional probabilities are interpretable and contain valuable economic information. Conversely, measures of declared confidence used in the literature are opaque, structurally biased, unstable, and more model-dependent, indicating that LLMs cannot assess their own confidence. Using conditional probabilities, we analyze LLM biases and provide insights into the internal mechanisms driving model decisions. Our results indicate that conditional probabilities provide a reliable and transparent reflection of LLM priors, particularly for economic applications.
