본 연구는 부도예측모형을 정보의 원천에 따라 회계모형, 시장모형, 그리고 헤저드모형으로 구분하고 각 모형의 부도예측력을 비교하였다. 회계모형은 분석방법에 따라 판별분석모형과 로짓분석모형으로 분류하였으며, 국내 기업에 적합한 변수를 새롭게 선정하여 변수의 계수를 재추정하였다. 시장모형으로 부도거리모형을 이용하였다. 회계정보와 시장정보를 통합하여 부도예측에 이용한 헤저드모형은 미국 기업에 적용하여 선정된 변수를 국내 기업에 그대로 적용하여 변수의 계수만을 재추정한 기존의 헤저드모형과 국내 기업에 적합하도록 모형을 수정한 새로운 헤저드모형을 이용하였다. 위 5개 모형의 부도예측력은 부도적중률, Receiver Operating Characteristic 곡선을 이용한 평가방법, 그리고 정보검증법으로 각각 평가되었다. 이 세 가지 평가방법에서 일관되게 국내 기업에 적합하도록 수정한 새로운 헤저드모형이 가장 부도예측력이 높게 나타났다. 그 다음으로는 미국 기업에 적용하여 선정된 변수를 국내 기업에 그대로 적용한 기존의 헤저드모형, 판별분석모형, 로짓분석모형, 그리고 부도거리모형순으로 부도예측력이 높게 나타났다. 본 연구의 결과는 다른 나라 기업에 적용된 부도예측모형을 그대로 사용하기보다는 국내 기업에 적합하도록 수정된 모형을 사용할 경우 부도예측의 정확성을 기할 수 있음을 시사한다고 볼 수 있다.
This paper evaluates the (out-of-sample) prediction performance of bankruptcy prediction models using Korean firms. Based on the source of information, we classify these models into accounting-based, market-based, and hazard categories. Note that hazard models are based on both accounting and market information. We consider five bankruptcy prediction models in this study; two accounting-based models, one market-based model, and two versions of a hazard model. One of the accounting-based models employs multivariate discriminant analysis (MDA), and the other employs logit analysis. The example of the first accounting-based model is the Altman (1968) Z-score model, and that of the second is the Ohlson (1980) O-score model. Most studies in the Korean literature use the accounting variables and coefficient estimates that Altman (1968) and Ohlson (1980) use for U.S. firms to predict bankruptcy, which may result in bias and inaccuracy because accounting variables may have different economic implications in different countries. Accordingly, in this study, we select new accounting variables that better fit Korean firms with respect to discriminant power and goodness-of-fit, and re-estimate the coefficients. For the market-based model, we use the KMV default-to-distance (DD) model. DD indicates the distance from the mean of the firm’s current asset value to its default point. The greater the DD, the smaller the probability of default. For the hazard model, we use the bankruptcy prediction model developed by Campbell, Hilscher, and Szilagyi (2008) (CHS), which uses both market and accounting information and has become popular in the finance literature. It has also been shown empirically that the CHS model is effective in predicting bankruptcy for firms in a variety of countries. We consider two versions of this hazard model. The first is one that adopts the same variables as those selected for U.S. firms in the CHS model and re-estimates them using Korean data. The second is a modified version of the CHS model. The variables are newly selected for Korean firms, and the coefficients are then re-estimated using data on these firms. The modification involves the addition of effective variables and exclusion of irrelevant variables based on empirical analyses. We estimate the five foregoing bankruptcy prediction models using data on all Korean firms in the seven years (in-sample period say, 2001~2007) prior to the one-year forecasting period (out-of-sample period say, 2008). By rolling over year by year, we repeat the in-sample estimation using the seven-year data prior to the one-year out-of-sample period. The sample period is 2001 to 2013. Thus, we obtain forecasting results from the five models for six years (i.e., 2008, 2009, 2010, 2011, 2012, and 2013). We then evaluate their bankruptcy prediction performance. Three methods are used to evaluate the models’ prediction accuracy for Korean firms: the hit ratio, receiver operating characteristic (ROC) curve, and information content test. The hit ratio is calculated as the ratio of the number of bankrupt firms in the portfolio to the number of all bankrupt firms in each forecasting year. Portfolios are formed by assigning all firms into one of ten decile portfolios based on the seven-year in-sample estimation results of each model. The hit ratio and ROC curve are traditional comparison methods that classify firms dichotomously (bankrupt or not). The information content test, in contrast, assesses whether different models convey information on actual bankruptcy. Of the five models evaluated, the hazard model modified for Korean firms performs best in predicting actual bankruptcies in the out-of-sample period with respect to all three bankruptcy prediction performance measures. The unmodified hazard model performs second best, and the DD model worst. There is little difference in prediction performance between the two accounting-based models, i.e., that using logit analysis and that using multivariate discriminant analysis, although their performance varies depending on the performance evaluation method used. This paper contributes to the literature in the following ways. First, by modifying the accounting and market variables used in the CHS model for Korean firms and re-estimating the model, we present a new hazard model that better fits Korean firms and outperforms other models in bankruptcy prediction accuracy. To the best of our knowledge, this is the first study to evaluate the performance of bankruptcy prediction models using Korean data. Second, this study is also the first in the Korean literature to include data over a period encompassing the 2008 global financial crisis. As there were many bankruptcies during that period, its inclusion enhances the reliability of our results. We suggest that our modified hazard model be used in future studies in both academia and industry.