변수간 공분산구조와 가중치를 고려한 사례기반추론의 적용: 기업부도예측 (Case-Based Reasoning Approaches by Considering Variable Covariance Structure and Variable Weight: Corporate Bankruptcy Prediction)

서지정보

· 발행기관 : 한국경영학회
· 수록지 정보 : 경영학연구 / 38권 / 5호 / 1165 ~ 1184페이지
· 저자명 : 홍효정, 조성빈

초록

본 연구는 기업부도예측을 위한 사례기반추론모형을 제시하고 전통적 통계모형과 인공지능모형과의 성과를 비교하고자 하였다. 기존의 연구에서는 모형 입력변수의 가중치를 차별화하는 것만 고려하였는데, 본 연구에서는 입력변수의 공분산구조를 추가적으로 고려하였다. 2001년에서 2003년에 걸쳐 수집된 중소규모의 제조업분야의 데이터를 대상으로 실험을 실시하였다. 사례기반추론 모형에서는 두 가지 실험요인(변수간 공분산구조를 고려함/고려하지 않음 × 변수가중치를 차별화함/차별화하지 않음)을 고려하여 총 네 가지 모형을 적용하였다. 가장 근접한 이웃에 포함되는 이웃의 수는 탐색을 통하여 15개로 결정하였다. 시뮬레이션기법을 적용하여 훈련용 데이터에서 가장 우수한 결과를 주는 모형을 선별한 후, 검증용 데이터에서 모형의 성과를 비교하였다. 실험 결과, 변수간 공분산구조와 가중치를 고려한 사례기반추론모형이 기존 모형에 비하여 높은 예측률을 나타내었다.

영어초록

Case-Based Reasoning (CBR) Approach has been used to predict future events by comparing past similar events. This paper proposes case-based reasoning models for corporate bankruptcy prediction problem and compares its performance with traditional statistical and artificial intelligence techniques. The Statistical techniques include multiple discriminant analysis and logistic regression, whereas artificial intelligence techniques include neural networks and decision trees. The purpose of this study is to develop a case-based reasoning model that reflects the covariance structure of variables and considers variable weights.
The literature review reveals that existing approaches in the area of CBR have mainly used the Euclidean distance for measuring dissimilarity between subjects and further differentiated variable weights for improving prediction accuracy. The distances between subjects are fairly different depending on whether the covariance structure of variables as well as variable weights are considered or not. Thus, we introduce four CBR models: Euclidean distance-based CBR without variable weight; Euclidean distance-based CBR with variable weight; Mahalanobis distance-based CBR without variable weight; Mahalanobis distance-based CBR with variable weight. We incorporate the covariance structure of input variables because the input variables of the model are commonly correlated with each other.
The data collected for analysis are the financial variables from the small-and medium-sized manufacturing firms in Korea during the fiscal year of 2001 to 2003. We selected 500 bankrupt firms and 500 non-bankrupt firms. After normalizing 132 variables, we obtained 15 variables by applying t-test, stepwise logistic regression, and decision trees induction. Then, by consulting former research, we finally selected seven variables, which include the various aspects of corporate activities such as profitability, liquidity, activity, stability, growth, and productivity.
The data are divided into the training data and the testing data set. The training data set is then divided into the training data set and the reference data set. Using 5,000 Monte Carlo simulations, the optimal weights of variables are determined while the nearest neighbors are searched from the training data set for predicting the bankruptcy of the reference data set.
Then, these weights are used to predict the bankruptcy of the testing data set.
The experiment results indicate that the CBR based on the variable covariance structure and variable weight produces a higher correct classification ratio than other CBR models and currentlyin- use approaches. The future study might improve practicality by applying different costs for the false positive prediction (predicting bankrupt for non-bankrupt firms) and the false negative prediction (predicting non-bankrupt for bankrupt firms).

참고자료

· 없음