Predicting Insolvency of Insurance Companies in Egyptian Market Using Bagging and Boosting Ensemble Techniques

Insolvency is a crucial problem for several insurance companies that suffer from it. This problem has direct or indirect effects on both the people working in the financial business and normal citizens. Thus, in insurance companies, the ability to predict insolvency is in great demand. There are several efforts proposed to predict insurance company insolvency using computer science methods (e.g., support vector machine and fuzzy systems). Each country has its own data patterns due to interior matters. Thus, insurance companies from different countries may have different data patterns. Consequently, the utilized predictive model should adapt to the dataset at hand. To our best knowledge, despite there are several efforts to build an insolvency predictive model, none of these efforts explored the Egyptian market. In addition, even the existing efforts did not utilize the ensemble learning methods in the insolvency prediction problem. In this context, we have two main contributions to this work. First, we proposed the first public access dataset of Egyptian insurance companies. The collected dataset was gathered from 11 Egyptian insurance companies during the years 1999 to 2019. The dataset consists of a set of 22 ratios (21 input features and one output feature), e.g., retention and investment yield alongside the solvency ration (i.e., the target feature). In the second contribution, we proposed exploring the performance of the ensemble learning methods to address the insolvency prediction problem. Thus, we proposed building several insolvency predictive models using ensemble learning and classic machine learning models. Next, the proposed models are evaluated on different accuracy metrics, e.g., Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The experimental results revealed that the ensemble learning-based models outperformed the classic machine learning-based models. Moreover, the correlation analysis between the utilized 22 financial ratios revealed that the most significant ratios, for the task of predicting the solvency ratio, are the technical provisions to shareholders’ funds, insurance companies’ debit balances to shareholders, and earnings after taxes to shareholders’ funds.