Friday, October 9, 2009: 9:40 AM
This paper presents the results of the first phase of a study comparing the ability of artificial intelligence models to predict bond rating changes. This first phase centers on the comparing effectiveness of genetic algorithms, in comparison to standard statistical models, to select the appropriate set of financial variables from a broad collection of variables that could be used to discriminate from firms that had a bond rating change from those corresponding firms that did not have a change in their bond ratings. We found a large number of studies in this field. Wilbur, 1977) examined corporate bond rating changes with respect to market efficiency. McAdams (1980); Soldofsky, Bhandari and Boe (1980) and Bhandari, Soldofsky and Boe (1983) used multiple discriminant (MDA) models to predict bond rating changes. Raman (1981, 1982) examined municipal bond rating changes. Copeland and Ingram (1982) also used MDA to predict municipal bond rate changes. Other authors - Metaew (1985) and Ibrahim, Metawae and Aly (1990) – have compared MDA models with another technique – statistical decomposition – for predicting bond rating changes. Other bond rating change studies that specifically compared alternative prediction techniques include Lewis, Patton and Green (1988) who compared statistical model vs analysts’ prediction of bond rating changes.
Data was collected on more than 900 bonds that had their Standard and Poor’s Corporation rating changed during the period 1997 to 2002. This period was selected because it reflected a number of different economic environments. We then matched this dataset with an equal number of corresponding firms which had the same initial bond rating but which did not change. The correspondence was based on the firms being in the same industry, having the same rating at the time of the change (the time frame was one month) and the same approximate asset size (within 20%). This relatively stringent set of criteria reduced the data set to 282 pairs of companies. For all pairs we had a broad range of financial measures that included ratios, changes in ratios and multi-year averages (72 variables in all). The goal is to reduce this set to predict both a bond change and the general direction of a movement from a particular bond rating to another bond rating. This research is somewhat similar to prior bankruptcy studies. The standard statistical methods used for variable reduction, in those studies, included multiple regression, factor analysis and discriminant analysis. We use these techniques and compare them with the results of a genetic algorithm. Genetic algorithms (or sometimes called evolutionary algorithms) were introduced by Holland (1975) as a method to perform randomized global search in a solution space. They offer tremendous potential for identifying patterns in complex environments. The second phase of this study uses both a neural network model and a multiple discriminant analysis to predict both a bond change and the general direction of a movement from a particular bond rating to another bond rating. The predictive variables will be the sets of financial ratios selected by the genetic algorithm and statistical tools.
Data was collected on more than 900 bonds that had their Standard and Poor’s Corporation rating changed during the period 1997 to 2002. This period was selected because it reflected a number of different economic environments. We then matched this dataset with an equal number of corresponding firms which had the same initial bond rating but which did not change. The correspondence was based on the firms being in the same industry, having the same rating at the time of the change (the time frame was one month) and the same approximate asset size (within 20%). This relatively stringent set of criteria reduced the data set to 282 pairs of companies. For all pairs we had a broad range of financial measures that included ratios, changes in ratios and multi-year averages (72 variables in all). The goal is to reduce this set to predict both a bond change and the general direction of a movement from a particular bond rating to another bond rating. This research is somewhat similar to prior bankruptcy studies. The standard statistical methods used for variable reduction, in those studies, included multiple regression, factor analysis and discriminant analysis. We use these techniques and compare them with the results of a genetic algorithm. Genetic algorithms (or sometimes called evolutionary algorithms) were introduced by Holland (1975) as a method to perform randomized global search in a solution space. They offer tremendous potential for identifying patterns in complex environments. The second phase of this study uses both a neural network model and a multiple discriminant analysis to predict both a bond change and the general direction of a movement from a particular bond rating to another bond rating. The predictive variables will be the sets of financial ratios selected by the genetic algorithm and statistical tools.