CNN-based deep learning approach for multiclass prediction of fin-tech startups

Fin-tech startups, while pivotal to financial innovation, face exceptionally high failure rates, underscoring the need for advanced predictive tools to guide investment and policy decisions. This study introduces a deep learning framework based on 1D Convolutional Neural Networks (CNNs) to classify fintech startup outcomes into four categories: Initial Public Offering (IPO), Merger & Acquisition (M&A), Closure, and Unicorn status using global Crunchbase data. CNN extracts local and hierarchical patterns from structured scalar attributes such as team composition, funding velocity, investor diversity, and company age. To address class imbalance, we adopt ADASYN oversampling. Compared to traditional models, XGBoost achieves 92.9% accuracy and 91.9% macro F1, while the CNN model achieves 90.8% accuracy and 89.6% macro F1, demonstrating competitive performance and improved sensitivity to rare outcomes such as IPOs and Unicorns. The results confirm tree-based dominance on structured data, although CNNs provide a viable deep-learning alternative for multiclass startup outcome prediction.