Doctoral Dissertations

Date of Award

8-1996

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Engineering Science

Major Professor

James A. Bontadelli

Committee Members

Robert E. Uhrig, Michael D. Vose, Elden L. Deporter

Abstract

This research discusses the development of an expert system methodology for identifying the fuzzy rules and membership functions used to predict the financial performance of small, technology-based companies. The training data consists of four years of financial information on 250 small, publicly traded, technology-based companies. A genetic algorithm determines the expert system fuzzy rules and membership functions using the training data. A fuzzy implication operator then determines the resultant output that represents the predicted financial performance. Additional companies are used to validate the fuzzy expert system model.

The training data consists of 114 financial and 31 economic independent variables for each company. Multiple linear regression analysis was used to determine the significant variables in predicting a company's financial performance as measured by the economic value added (EVA). The significant independent variables are then used by a genetic algorithm to derive the "best fit" fuzzy rules and membership functions to develop the expert system model. This research found that none of the 31 external economic variables significantly contributed to predicting a company's EVA.

The number of fuzzy rules identified using a genetic algorithm strongly corresponds to the number of data clusters found using k-mean cluster analysis. Additional information available in the training data for learning new rules decrease significantly beyond a specific number of data clusters. As a result, this research suggests that cluster analysis may be used to identify the number of fuzzy rules potentially available in a training data set.

The fuzzy expert system model with 20 or more rules outperformed the "best fit" ten variable multiple linear regression model in predicting a company's EVA. After reducing the number of independent variables to include only the five most significant independent variables that contribute to improving the regression model's coefficient of determination, the fuzzy expert system model produced additional prediction improvement. This improvement is partially due to the reduced search space resulting in a higher probability that the genetic algorithm will identify a "better fit" fuzzy rule and membership function combination.

This research found that a fuzzy expert system model with fewer significant independent variables was less sensitive to system noise and more responsive to general trends. Therefore, the fuzzy expert system that included independent variables which contribute measurably to increasing the regression model's coefficient of determination, significantly outperformed the multiple linear regression model, with the same variables, in predicting a company's future financial performance.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS