Using GPT to Measure Business Complexity

ABSTRACT Business complexity involves important tradeoffs for managers and investors, but empirical evidence is limited by measurement issues. We construct and validate a measure of business complexity using a GPT model fine-tuned on narrative disclosures and inline XBRL tags. We first show that our measure is associated with slower price formation in capital markets, consistent with complexity increasing processing costs. Next, we apply our measure to study the complexity of debt, an economically important topic that encompasses a wide range of features. The results show that nonstandard debt features such as call and convertibility provisions underlie debt complexity. We also find that debt complexity correlates with more persistent interest expense and better performance when lending conditions worsen, suggesting it is in part an adaptive response to manage financial risk. Overall, our study underscores the tradeoffs of business complexity and provides a flexible measure of complexity for future research. Data Availability: Contact authors for data, model weights, and measure. JEL Classifications: D82; D83; G14; G30; M40; M41.