Early-Stage Startup Success Prediction using Ensemble Machine Learning for Venture Capital Investment Decisions

Predicting early-stage startup success remains a major challenge for venture capital due to high uncertainty and extreme class imbalance. This study develops a large-scale machine learning framework using Crunchbase data (462,651 startups, 186 features) to screen and rank investment opportunities. We compare Random Forest, SVM, Logistic Regression, XGBoost, Gradient Boosting, and a voting ensemble, and identify three robust predictors across models: number of funding rounds, network relationships, and total capital raised. A tuned Gradient Boosting model achieves 89.3% accuracy on a held-out test set; however, given a 2.4% base success rate, evaluation focuses on precision–recall behaviour rather than accuracy alone. Deployment on an independent dataset illustrates practical screening utility, while highlighting the key limitation of the absence of temporal validation for true forecasting. The framework is intended to support, not replace, expert venture capital decision-making and provides a scalable, interpretable approach to startup evaluation.