Abstract:

We study how early-stage investors evaluate potential investments by using large language models (LLMs) to analyze 6,800 expert consultation calls. Not only do call volume and overall sentiment predict outcomes, but the specific content of discussions provides significant additional predictive power. Our topic-specific sentiment analysis shows that positive signals about technology integration and customer acquisition are associated with 15% and 16% higher deal likelihood, respectively. We find that the information content of the calls is particularly valuable for younger firms with limited track records, where information asymmetries are most severe. Our findings provide the first systematic evidence of how investors gather and process information in the absence of traditional financial metrics, and suggest some misalignment between topics that investors frequently discuss and those that best predict deal outcomes. Methodologically, we demonstrate the potential of LLMs to extract nuanced insights from complex qualitative data.

Authors:

Victor Lyonnet, University of Michigan
Amin Shams, Ohio State University
Shaojun Zhang, Ohio State University