Machine Learning with Synthetic Data for Predicting High Growth Companies

Sammanfattning

Identifying high-growth firms is a critical challenge in finance and strategic investment, particularly within the context of mergers and acquisitions (M&A). This thesis explores the application of machine learning and synthetic data generation to improve early-stage detection and prediction of high-growth companies, using three-year samples of standardized financial statement data for private limited companies in Sweden, Norway, and Finland. To address the rarity of high-growth firms, which introduces severe class imbalance, we employ Conditional Tabular GANs (CTGAN), a deep generative model tailored to tabular data, to generate synthetic minority class samples. These are used to augment the real training data, with the goal of improving classification performance. CatBoost, a gradient boosting algorithm specifically optimized for categorical and tabular features, is used as the primary model for binary classification. Performance is evaluated using ROC-AUC, PR-AUC, and F1-score metrics, while SHAP values are used to interpret feature importance. The results show generally good classifying accuracy considering the class imbalance. It was also seen that augmenting real data with synthetic high-growth samples offers modest but consistent performance improvements in certain specific configurations but still has some limitations in other contexts. SHAP analysis reveals that company size, asset growth, and age are among the most predictive features for identifying high-growth firms. These findings suggest that synthetic data generation, in some cases, can support machine learning models in early identification of high-growth companies.