On multiple‐class prediction of issuer credit ratings

Abstract

For multiple-class prediction, a frequently used approach is based on ordered probit model. We show that this approach is not optimal in the sense that it is not designed to minimize the error rate of the prediction. Based upon the works by Altman (J. two-class prediction, we propose a modified ordered probit model. The modified approach depends on an optimal cutoff value and can be easily applied in applications. An empirical study is used to demonstrate that the prediction accuracy rate of the modified classifier is better than that obtained from usual ordered probit model. In addition, we also show that not only the usual accounting variables are useful for predicting issuer credit ratings, market-driven variables and industry effects are also important determinants.Currently, there are many widely recognized rating agencies, such as Moody's Investors Service, Standard and Poor's Ratings Services (S&P's), etc. They routinely provide credit ratings for bonds and companies.This study focuses on the S&P's long-term issuer credit rating (LTR). According to the definition given by S&P's, the LTR focuses on the obligor's capacity and willingness to meet its long-term financial commitments. To determine the LTR of a company, S&P's examines a profile called the corporate rating analysis and methodology profile. The profile contains two types of information of the company. The first type of information is available publicly, for example, public financial data. The second type of information is collected through a proprietary process from industry characteristics, competitive bargain positions, interviews with management team, etc. Thus, S&P's rating procedure is in part common knowledge. However, the major determinants of S&P's LTR are basically not clear. In this paper, we propose to first identify important predictors of S&P's LTR, selected from publicly available market data, accounting data, and industry classification.Based on the Compustat North America (COMPUSTAT) database, there were 8039 companies in the year 2005 having stock traded on the New York Stock Exchange, American Stock Exchange, or NASDAQ. However, among those 8039 companies, there were only 20.46% (1645) companies having S&P's LTRs. This means that most of companies do not have their S&P's LTRs. In this paper, our next focus is to forecast ratings of those companies 'without' S&P's LTRs. Pettit et al. [1] reported that the new faces in the pool of companies with S&P's LTRs have lower rating category on the average. Blume et al. [2]also found similar results that bond ratings have declined, but the decline could be due to the use of more stringent rating standards in assigning ratings. On the other hand, we do not pursue the issue of rating forecast for companies 'with' S&P's LTRs for two reasons. First, if a company is once rated by S&P's, then it will be continuously rated unless a special event happens to the company, for example, bankruptcy. Second, the continuously rated companies have relatively unchanged rating categories in general [1]. Thus...

Predicting issuer credit ratings using a semiparametric method

“…This indicates that marketdriven variables and industry effects are also important to determine S&P's LTRs. Our variable selection result coincides with that obtained by Hwang (2008) for predicting ratings in year 2005.…”

“…To apply OLPM and OSPM to predict S&P's LTRs, the twenty-four potential predictors in Hwang et al (2008) for studying important predictors of S&P's LTRs in year 2005 were considered in our data analysis section. These variables include four market-driven variables (Shumway, 2001;Bharath and Shumway, 2008), nineteen accounting variables (Altman, 1968;Poon, 2003;Pettit et al, 2004), and industry effects (Chava and Jarrow, 2004;Pettit et al, 2004).…”

“…Our sample consisted of 779 companies receiving S&P's LTRs in April 2007 and having complete values of the twenty-four potential predictors. The sample was further divided into the estimation sample and holdout sample based on the longevity of S&P's LTR (Hwang et al, 2008). 2 According to S&P's Research Insight North America Data Guide (2004, p. 54), S&P's began to use the term LTR on September 1, 1998.…”

Predicting issuer credit ratings using generalized estimating equations

“…One is the static model using cross-sectional data of sampled companies. The static models include the ordered probit model (Kaplan and Urwitz 1979, Ederington 1985, Gentry et al 1988, Hwang et al 2009), multiple regression analysis (Horrigan 1966, Pouge and Soldofsky 1969, West 1970, multiple discriminant analysis (Pinches andMingo 1973, Altman andKatz 1976), ordered and unordered logit models (Ederington 1985), etc. A detailed introduction of these models can be found in the monograph of Altman et al (1981).…”

A hybrid KMV model, random forests and rough set theory approach for credit rating

“…The former include linear regression [26,47,54], linear multivariate discriminant analysis (MDA) [2,46], probit regression [28,32], logit analysis [31] and multidimensional scaling [40]. The latter consists of neural networks [15,22,35], case-base reasoning [34,50] and support vector machines (SVM) [3,8,27,33].…”