Expert systems in insurance underwriting: model development and application

The paper examines the work of Dr. Gary A. Wicklund and doctoral student Ms. Roberta M. Roth of the University of Iowa in the development and implementation of an insurance underwriting expert system for a Midwestern insurance company. The underwriting function reviews applicant data for determination of insurability. The feasibility of utilizing an expert system in this area was examined and determined to be practical, and development of a prototype was initiated. The authors began a broad search for and examination of available expert system tools for a personal computer environment. The review of six expert system software packages narrowed the field to two viable candidates. The Ml package from Teknowledge was selected for the first prototyping stage. The prototype was quickly developed with a very limited set of rules to provide a simple demonstration of the applicability of an expert system in this area. The demonstration of the expert system concepts through the prototype to upper management and underwriting management at a local insurance company was well accepted and authorization to build the expert system was given. The host mainframe at the sponsoring insurance company proved incompatible with the M1 software, which was replaced by RuleMaster from Radian. The prototyping work was transferred to RuleMaster, which proved to be beneficial beyond the compatibility concerns. RuleMaster is a modular system as opposed to the inference network approach used in M1, which more closely resembles LISP or PROLOG in structure. With the modular structure, knowledge and rule base building and modifications proved much easier from a programming standpoint. The interviewing of the domain expert was very valuable for future reference in the next domain to be prototyped. Because of the inexperience of the parties, the interviews were unstructured and the knowledge acquisition was a laborious task for the knowledge engineer and the expert. The next interviews were videotaped and replayed later to allow the knowledge engineer closer scrutiny of the responses and adequate time for formulation of additional inquiries. The domain expert and knowledge engineer reviewed sample applicant data to determine key fields and values in those fields which triggered thought processes in the domain expert. By using this approach, the programming of the modular system was simplified, as the software also uses key fields and values which interact with rules for treatment of possible values and provides classification of these values. This also aided the domain expert in understanding how his reactions would be utilized and assisted him in his responses regarding his thought processes. After the initial prototyping and knowledge and rule base building, the system began its refinement stage. During this period the system ran identical data with the domain expert, for comparison with the conclusions which were reached. Discrepancies between the system and the domain expert identified shortcomings and areas for refinement. RuleMaster again proved beneficial in that it provides reasoning for the conclusions reached, which could then be more closely examined by the domain expert and knowledge engineer for modification. Another programming asset of RuleMaster is its ability to induce rules from examples and their treatment which are entered in a simple table by the knowledge engineer. These tables can be easily modified and all changes are immediately reflected in RuleMaster. The sponsoring insurance company also requested that the conclusions include a conclusion strength factor (in this case numeric) which would be highly recognizable to the user as an additional supporting measurement of the conclusion reached. The sponsor will soon begin to run the system parallel with its entire underwriting staff for further analysis and as the first step toward implementation.