Life Insurer Insolvency Warning System: A Neural Network Method

First Executive Corporation’s bankruptcy in 1991 made the insurance policyholders, investors and regulators to concern about the solvency of life insurer in US market. As First Executive Corporation had more than 15 billion dollars in assets, its failure provoked increasing debate over the prediction system of financial distress in insurance industry, the Insurance Regulatory Information System (IRIS). From 1970s, National Association of Insurance Commissioners (NAIC) has developed IRIS which serves as a baseline solvency screening system. IRIS helps regulators prioritize insurers for detailed financial analysis and allocate their resources according to need by producing 12 financial ratios for life insurers and 11 for property-casualty insurers (Atchinson 1996). When an insurer has four or more of those ratios beyond the specified range, the state regulators and NAIC will classify it as priority firm and pay immediate attention. Two years after failure of First Executive, NAIC adopted the Financial Analysis and Solvency Tracking System (FAST) as an expansion of IRIS. FAST assigns different points for different ranges of 29 financial ratio results. The higher the total score is, the higher attention the state regulators and NAIC pay for. Essentially, IRIS is the implementation of univariate approach of discriminant analysis first proposed by Beaver (1966), and FAST is the implementation of multivariate analysis first proposed by Altman (1968). From 1980s, the logistic analysis has replaced multivariate analysis as a most used statistical method for insolvency prediction purposes, since multivariate analysis suffers from its normality assumption. Except the prediction systems used by NAIC, financial academy and practitioner proposed a number of other methods to predict business insolvency, ranging from nonfinancial industries to financial industries. In the insurance industry, most previous studies have been made on the property & casualty insurers, but very limited of them have been applied to life & health sector. More important, most approaches applied to insurance companies are statistical methods such as discriminant or logit analysis. Both of them suffer from their statistical assumptions about the observed information which is a set of financial ratios in most cases. Recently, the integration of artificial intelligence with the financial economics proposed neural network as a promising alternative for classifying the insolvency financial institutions (Brockett 1994; Tam 1990; Salchengerger 1992; Wilson 1994). In this paper, we used a neural network method in life insurer insolvency prediction. Following (Brockett 1994), we constructed a feed-forward backpropagation three-layer artificial neural network, but we applied this method to life insurers listed in NAIC InfoPro data tape. We also compared the predicating ability of stepwise multivariate linear discriminant analysis, the stepwise multivariate logit analysis, and neural network. Using matched data of solvent and insolvent life insurers in the time period from 1993 to 2001, we showed that a neural network model with feed forward, back propagation algorithm can provide satisfactory insolvency predictive ability compared to the traditional methods. Although the neural network model is promising in the prediction of business insolvency, a lot of jobs are still needed to do in the future. At the end of the paper, we further investigate the convergence condition and the proper repeating time.