A Computational Model of Loan Loss Judgments
暂无分享,去创建一个
Key Words: Computational model, Judgment expertise, Knowledge-based systems. Data Availability: Details on the data are available from the first author. The viability and profitability of a financial institution is directly affected by the financial integrity of its loan portfolio. External auditors evaluate the collectibility of loans to estimate the uncollectible portion of the portfolio, if any, with direct, and potentially large, effects on the client's financial statements. Loan evaluations can be difficult because the judgment is not well-structured given "textbook" knowledge, repayment circumstances are usually highly contextual, and considerable qualitative and quantitative information may be relevant (Hale 1983). Surprisingly, given the importance of the judgment for auditors, little is known about how experienced auditors evaluate loan collectibility. For example: ... the auditor's evaluation of a financial institution's loan portfolio value (net) is highly judgmental in nature; more of an art than a science. Very little guidance exists on how to perform individual loan evaluations. (Ribar et al. 1990, 1). The current research addresses this lack of knowledge. A model is presented of the process of deciding what portion, if any, of a domestic commercial loan is now uncollectible. There are several reasons for this research. First, the model indicates how the judgment is performed, including the contextual richness necessary for each borrower situation. The model therefore can be evaluated and extended by both academics and practitioners. Second, the model may be helpful for staff training of auditors. Third, a basis is provided for research on when decision support can be most productively applied to aid auditors in evaluating loans (e.g., Wright 1995). Fourth, availability of such a model permits integration with, and provides guidance for, results from other research paradigms such as experiments (Libby 1975; Wright 1996) and econometric models (Dietrich and Kaplan 1982). Fifth, the model can be used as the basis of a knowledge-based, "expert" system. The model reported here is an extension and elaboration of an early version of the model entitled "CFILE" (Kelly et al. 1986).(1) The current model includes more knowledge and reasoning. For example, now the control structure (figure 1) has more depth and the knowledge in the subgoal structure (figure 2) has been extended and refined significantly, including more contextual reasoning. The depth and contextual appropriateness of the three loan evaluation examples reported in the paper illustrate the considerable enhancements of the CFILE model. CFILE is a prototype; "Loan Auditor" is a "finished" model. [Figures 1-2 ILLUSTRATION OMITTED] The paper is organized as follows. The next section presents the computational modeling methodology. Section three describes the research process. Section four presents details of the model, including specifics of knowledge use and reasoning strategies, and representation thereof, as well as three detailed process descriptions. Model validation is reported in section five. A summary of the model and its limitations, as well as ideas for future research, constitute the final section. COMPUTATIONAL MODELING OF LOAN JUDGMENTS Computational modeling produces a theory of the process of performing a complex, semi-structured judgment task that requires considerable task-specific knowledge for high level performance (Bailey et al. 1988, 13-27; Biggs 1991; Peters 1990, 1993). Biggs (1991, 1) describes computational modeling as follows: This type of modeling involves developing a detailed description of the knowledge and reasoning processes used by a human to per form a task. This description represents a theory or model of human task performance which is then implemented in a computer program that actually performs the task according to the specifications of the descriptive theory. …