딥러닝 알고리즘에 기반한 기업부도 예측

Abstract

기업의 부도는 국가경제에 막대한 손실을 입히며, 해당기업의 이해관계자들 모두에게 경제적 손실을 초래하고 사회적 부를 감소시킨다. 따라서 기업의 부도를 좀 더 정확하게 예측하는 것은 사회적·경제적 측면에서 매우 중요한 연구라 할 수 있다. 이에 최근 이미지 인식, 음성 인식, 자연어 처리 등 여러 분야에서 우수한 예측력을 보여주고 있는 딥러닝(Deep Learning)을 기업부도예측에 이용하고자 하며, 본 논문에서는 기업부도예측 방법으로 여러 딥러닝 알고리즘 중 DBN(Deep Belief Network)을 제안한다. 기존에 사용되던 분석기법 대비 우수성을 확인하기 위해 최근까지 기업부도예측에서 연구되고 있는 SVM(Support Vector Machine)과 비교하고자 하였으며, 1999년부터 2015년 사이에 국내 코스닥·코스피에 상장된 비금융업의 기업데이터를 이용하였다. 건실기업의 수는 1669개, 부도기업의 수는 495개이며, 한국은행의 기업경영분석에서 소개된 재무비율 변수를 이용하여 분석을 진행하였다. 분석결과 DBN이 SVM보다 여러 평가척도에서 더 좋은 성능을 보였다. 특히 시험데이터에 대해 부도기업을 부도기업으로 예측하는 민감도에서 5%이상의 더 뛰어난 성능을 보였으며, 이에 기업부도예측분야에 딥러닝의 적용가능성을 확인해 볼 수 있었다.| Corporate bankruptcy causes serious damage to national economy, brings about economic losses to all the concerned of the company, and reduces social wealth. Therefore, forecasting corporate bankruptcy more accurately is very important in the social·economical aspect. In this thesis, Deep Learning method, which shows excellent predictive power recently in many fields, such as image recognition, voice recognition and natural language process, is used to forecast corporate bankruptcy, and specifically it suggests DBN(Deep Belief Network) among many deep learning algorithms. In order to verify performance of the method comparing with existing analysis technique, this thesis tried to make a comparison with SVM(Support Vector Machine) which is studied for corporate bankruptcy forecast method until a recent date, with corporate data of non-financial business listed on domestic KOSDAQ·KOSPI from 1999 to 2015. The number of healthy companies was 1669 and the number of bankrupt companies was 495; analysis was performed using financial ratio variables introduced in `Business Management Analysis' of the Bank of Korea. As a results, DBN exhibited better performance in various rating scales than SVM. In particular, it showed more than 5% better performance in sensitivity of predicting a bankrupt company as a bankrupt company based on test data, and it was able to confirm applicability of deep learning in the field of corporate bankruptcy forecast.; Corporate bankruptcy causes serious damage to national economy, brings about economic losses to all the concerned of the company, and reduces social wealth. Therefore, forecasting corporate bankruptcy more accurately is very important in the social·economical aspect. In this thesis, Deep Learning method, which shows excellent predictive power recently in many fields, such as image recognition, voice recognition and natural language process, is used to forecast corporate bankruptcy, and specifically it suggests DBN(Deep Belief Network) among many deep learning algorithms. In order to verify performance of the method comparing with existing analysis technique, this thesis tried to make a comparison with SVM(Support Vector Machine) which is studied for corporate bankruptcy forecast method until a recent date, with corporate data of non-financial business listed on domestic KOSDAQ·KOSPI from 1999 to 2015. The number of healthy companies was 1669 and the number of bankrupt companies was 495