Evaluating Founder Signals Under Uncertainty: A Predictive-Modeling Frame-work Using LLM Policy Induction and Gradient Boosting

Sammanfattning

This thesis investigates whether founder human- and social-capital signals retain predictive value across time when forecasting early-stage venture funding outcomes. Addressing recent calls for rep-lication and temporal validation in signaling research, the study develops a predictive-modeling framework that combines a novel LLM-based policy-induction model with established machine-learning methods. Using a dataset of more than 50,000 startup founders in Europe and North Amer-ica, the models are trained on the five yearly cohorts preceding each prediction year (2012-2019) and then evaluated on later, strictly unseen cohorts (2018-2022) to test temporal generalizability under rigorous out-of-sample conditions. The results demonstrate strong predictive stability across cohorts and reveal three consistent patterns. First, prior entrepreneurial experience remains a persis-tent and influential predictor of high funding attainment. Second, Big Tech experience functions as a rare but highly rewarded signal, consistently outperforming academic credentials such as PhD train-ing or research roles. Third, academic credentials contribute positively only when embedded within strong execution-oriented profiles, and never as standalone drivers of prediction. They function as weak, conditionally relevant signals that the models treat as secondary to operational experience ra-ther than as sources of genuine configurational synergy. Methodologically, the thesis displays how predictive modeling and interpretable LLM-based reasoning can be combined to evaluate theoretical mechanisms with temporal rigor.