Warren Buffett once advised that you should never invest in a business you cannot understand. Yet, that principle has not deterred many investors from backing companies whose inner workings remain opaque. In increasingly complex markets, where financial structures and disclosures have grown more intricate, the gap between what investors are told and what they truly comprehend continues to widen. This disconnect has created a demand for better tools that can help interpret and assess the underlying complexity of modern corporations.
New research from the McCombs School of Business offers a promising step in that direction. The study introduces what may be the most precise and comprehensive method yet for measuring business complexity, providing investors and analysts with a clearer lens through which to evaluate corporate structures. By refining how complexity is defined and measured, the research addresses a longstanding challenge in financial analysis: capturing the true depth of a company’s operations beyond surface-level indicators.
The tool, developed by Sara Toynbee, an associate professor of accounting, leverages artificial intelligence to simplify this traditionally difficult task. Rather than relying on crude proxies such as company size or the number of operating segments, the model evaluates complexity from the perspective of an external observer. It asks a straightforward but powerful question: how difficult is it to understand a firm’s financial position and performance based on the information disclosed in its reports? This reframing shifts the focus from structural attributes to interpretive difficulty, offering a more nuanced assessment.
Historically, measuring complexity has been challenging precisely because complexity itself is multifaceted. Conventional metrics often fail to capture deeper layers related to financial instruments, risk exposures, or reporting practices. Toynbee’s approach recognises that complexity does not arise from a single source but can vary widely across firms. Her model addresses this by incorporating 29 distinct categories, including debt, equity, derivatives and hedging, taxation, revenue recognition, and executive compensation. By doing so, it creates a multidimensional profile that reflects how complexity manifests across different aspects of a business.
To build the model, Toynbee collaborated with Darren Bernard, Elizabeth Blankespoor, and Ties de Kok from the University of Washington. Together, they trained a large language model based on Llama 3 using 200,000 sentences drawn from financial statement footnotes. These sentences included embedded iXBRL tags, which provide machine-readable labels describing the meaning of numerical values. By learning to predict what each number represents based on its context, the model effectively mimics the interpretive process of a skilled human reader, but at a vastly greater scale.
After training, the model was applied to more than eight million individual numerical disclosures across over 50,000 corporate reports spanning 2016 to 2024. Each number was assigned a complexity score based on how difficult it was for the model to classify accurately. The less confident the model was in its classification, the higher the inferred complexity. This approach transforms complexity into a measurable, data-driven concept, enabling systematic comparisons across firms and over time. It also reveals patterns that would be nearly impossible to detect through manual analysis alone.
The findings suggest that complexity carries both costs and benefits. On the one hand, higher complexity appears to slow the market’s response to financial disclosures, with stock prices taking longer to adjust as investors process more intricate information fully. On the other hand, complexity can also serve a strategic purpose. In areas such as debt structuring, more complex arrangements, including instruments with non-standard terms like convertibility into equity, can help firms manage risk and stabilise financial outcomes. The research highlights that while excessive complexity may obscure understanding, certain forms can enhance resilience, offering a more balanced perspective on its role in modern business.
More information: Darren Bernard et al, Using GPT to Measure Business Complexity, The Accounting Review. DOI: 10.2308/TAR-2023-0716
Journal information: The Accounting Review Provided by University of Texas at Austin