Forecasting firm growth resumption post-stagnation

Abstract

Our study forecasts the likelihood of firms resuming growth after periods of stagnation or declining sales. We employ machine learning methods, including Random Forest, LightGBM, and CatBoost, alongside logistic regression models. To address class imbalance, we incorporate oversampling techniques such as SMOTE, ADASYN, and SMOTEENN. We focus on two key indicators—Precision (predictive accuracy) and Recall (completeness of prediction)—to meet the needs of different investor groups. The performance of our models is evaluated using metrics such as accuracy, precision, recall, F-score, and RocAUC, with Venkatraman's test applied for model comparison. Our key findings reveal that CatBoost achieves a predictive accuracy of 65–67 %, significantly outperforming random firm selection, which yields only 13–17 % accuracy. The combination of the CatBoost method with the SMOTEENN technique notably enhances Recall values, reaching 58–63 %, a critical metric for large investors and policymakers. Our study offers a methodological approach to better understand and forecast the trajectories of firms engaged in open innovation.