Learning from Yesterday: Predicting early-stage startup success for accelerators through content and cohort dynamics

As the demand for seed accelerators grows, so does the complexity of their evaluations of numerous startup applications. This paper introduces a novel two-phase data-driven framework for startup performance prediction. Phase 1 extracts founding team-level and venture-level features applicable to early-stage startups for success prediction. Phase 2 further engineers cohort-level features to predict the success of accelerator-admitted startups. We demonstrate the utility of our framework by leveraging machine learning methods coupled with real-world data of 35,647 startups (accelerator intakes: 763). We achieve high predictive accuracy and produce explainable results. We make methodological contributions to startup competitor detection and industry categorization. The key insight of our study is that member success largely depends on cohort-level features such as shared industries with different members and industry similarity to the accelerator's past portfolio.