A Neural Network Method for Obtaining an Early Warning of Insurer Insolvency

Introduction The definition and measurement of business risk has been a central theme of financial and actuarial literature for years. Although the works of Botch (1970), Bierman (1960), Tinsley (1970), and Quirk (1961) dealt with the issue of corporate failure, their models did not lend themselves to empirical testing. Additionally, while the work by Altman (1968), Williams and Goodman (1971), Sinkey (1975), and Altman, Haldeman, and Narayanan (1977) attempted to predict bankruptcy by using discriminant analysis, their approaches were static in nature. Thus, the dynamics of a firm's operations and the changing economic environment were not included in their analyses. Santomero and Vinso (1977), and Vinso (1979) used actuarial ruin theory models to incorporate the dynamic aspects of the cash flow process for assessing the likelihood of bank insolvency. However, there appear to be mathematical problems with their development, making their results difficult to implement in practice. Insolvency within the insurance industry has become a major issue of public debate and concern, and the identification of potentially troubled firms has become a major regulatory research objective. Previous research on the topic of insurer insolvency prediction includes Ambrose and Seward (1988), BarNiv and Hershbarger (1990), BarNiv and MacDonald (1992), Barnoff (1993), and Harrington and Nelson (1986). BarNiv and MacDonald provide a particularly good review of the previous research techniques and results on rating and monitoring insolvency risk for insurers and can be consulted for further background on alternative approaches. The research presented in this article aims to construct an early warning system for regulatory use in insolvency prediction. The approach we utilize is based upon modern methods in artificial intelligence (in particular, a neural network model), and uses financial and other insurer operations data such as those available in the annual statements filed with the National Association of Insurance Commissioners. We also compare the insolvency prediction results using the neural network methodology with those obtained via discriminant analysis, A. M. Best ratings, and the National Association of Insurance Commissioners' Insurance Regulatory Information System ratings. Situational Overview In the context of warning of pending insurer insolvency, the regulator has several sources of information. For example, there are reporting and rating services such as the A. M. Best Company, which rates 3,000 property-liability and life health insurers. However, many of the insurers of interest to state regulators are not rated by Best's or by other rating services (e.g., Moody's or Standard and Poor's). In addition, the National Association of Insurance Commissioners (NAIC) has developed a system called the Insurance Regulatory Information System (IRIS). This system was designed to provide an early warning system for insurer insolvency based upon financial ratios derived from the regulatory annual statement. The IRIS system identifies insurers for further regulatory evaluation if four of the eleven (or twelve, in the case of life insurers) computed financial ratios for a particular company lie outside a given range of values. IRIS uses univariate tests, and the acceptable range of values is determined such that, for any given univariate ratio measure, only approximately 15 percent of all firms have results outside of the particular specified range. The adequacy of IRIS for predicting troubled insurers has been investigated empirically and found not to be strongly predictive. For example, one small-scale comparison study using only five of the IRIS ratio variables from the NAIC data has shown that by using statistical methods it is possible to obtain substantial improvements over the IRIS insolvency prediction rates (cf. Barrese, 1990). More recently, the NAIC has implemented a supplementary system based on a number of additional financial ratios. …