New TRUST initiative to bring transparent AI to the corporate bond market

Two Virginia Tech researchers have launched Transparent Risk Understanding through Structured Tone-analysis (TRUST), a multidisciplinary initiative that uses artificial intelligence (AI) to transform how investors, analysts, and regulators assess corporate bond risk. 

The initiative aims to identify early warning signals in corporate communications, helping market participants make more informed decisions.  

Initiative co-leaders Chang-Tien Lu, professor of computer science and associate director of the Sanghani Center for Artificial Intelligence and Data Analytics, and Sattar Mansi, the Wells Fargo Professor of Financial Risk Management in the Pamplin College of Business, are combining their respective cutting-edge machine learning and financial domain expertise to deliver tools that are not only accurate, but also understandable and verifiable. Both professors are based in the greater Washington, D.C., area.

Computer science Ph.D. students Shengkun Wang and Linhan Wang, advised by Lu, also work on the TRUST project.

^{A replacement for backward-looking tools} 

Today, pension funds and fixed-income investors still rely heavily on balance sheet indicators, leverage ratios, profitability metrics, and broad macroeconomic factors. 

“These tools are backward-looking and often miss early warning signals embedded in earnings calls, management commentary, and regulatory filings until markets react through widening credit spreads or sudden downgrades,” said Lu, who is also a faculty member with the Institute for Advanced Computing.

Two historical examples illustrate this point: General Electric and Enron. General Electric executives signaled cash flow and restructuring challenges months before the company’s 2018 credit downgrades. In the case of Enron, disclosures grew increasingly complex and defensive well ahead of its 2001 collapse, subtly indicating deeper issues that were difficult to detect systematically.

The TRUST system is designed to change this paradigm. 

TRUST employs an advanced agentic AI pipeline that integrates FinBERT, a pre-trained model for analyzing financial text, large language models, and an optical character recognition-enhanced vision–language module capable of parsing scanned tables and embedded financial graphics. 

“Using real-time, AI-driven insights add a new dimension to credit markets,” said Lu. “TRUST produces an integrated workflow, continuously analyzing earnings calls, annual reports, regulatory filings, and financial data to identify shifts in managerial tone, narrative consistency, and financial alignment. It produces transparent ‘tone cards’ that not only highlight an emerging risk but are designed to help analysts understand the reasoning process and the documentation that supports each finding."

^{Improved decision-making and market efficiency}

“Integrating cutting-edge machine learning with financial expertise allows us to deliver tools that enhance transparency, repeatability, and practitioner review, each essential for high-stakes financial decision-making,” Mansi said. “More timely risk assessment, smoother market adjustments, and bond pricing that reflects not only financial metrics but also the quality and credibility of corporate disclosures, can ultimately improve market efficiency.”

The final goal, said the researchers, is to build a trustworthy system that regulators, investors, and researchers can actually audit and rely on.

They plan to release a public replication package, including sample datasets, code modules, and documentation, to encourage collaboration and support robust scientific review.

“With growing interest in trustworthy AI across industry and academia, we hope to contribute to Virginia Tech’s position as a leading contributor to next-generation analytics for corporate credit markets,” Lu said.