以類神經網路建構階段式企業財務危機預警模型

Using Neural Networks to Predicate Financial Distress : A Multi-state Model

陳振嘉(Jeng-Chia Chen)

指導教授 ： 許通安

中原大學/商學院/資訊管理學系/碩士(2006年)

https://doi.org/10.6840/CYCU.2006.00423

摘要

提前預測企業財務危機可為債權人與投資者節省因跳票、倒閉等之金錢損失，以往的研究往往採用二元分類（Binary-state）模型，忽略企業發生財務危機為一連續性的過程。本研究嘗試以多重分類（Multi-state）的方式，分析企業的真正財務狀態，使用倒傳遞類神經網路與次序性Logistic迴歸模型相比較。分析的變數除了傳統財務比率外，也加入現金流量、股權結構與會計師資訊等非財務變數，並分別採用「因素分析」與「逐步迴歸」篩減變數，針對模型的效益評估特別著重模型的「穩定性」與「分類誤差」，選用五年的資料期間每年建立不同的預警模型，藉此評估於期間內相對穩定的預警模型，在分類誤差方面，則針對多重分類的特性以型一誤差比率與預期分類誤差成本（Expected Cost of Misclassification）作衡量。

本研究總共建立45個財務預警模型，經過準確率、型一誤差、ECM的評估後發現「倒傳遞類神經網路」在多重分類上得效益遠較「次序性logistic迴歸模型」為佳，使用「因素分析」的模型具有穩定性stability高與誤差小的優點，篩選出的變數則以會計師資訊、獲利能力、財務結構指標佔多數，以「等比例樣本」為訓練樣本的類神經網路模型也在穩定性上勝過「實際樣本比例」模型，而結合「因素分析」與「等比例樣本」建構的倒傳遞類神經網路模型更是為9種模型較為優秀者。

關鍵字

多重分類 ； 財務危機預警模型 ； 倒傳遞類神經 ； 次序性logistic迴歸

並列摘要

Predicting corporate financial distress helps to debtors and investors to presents economic loss from corporate bankruptcy. In the past academic research, Binary-State Model is most used to forecast the crisis, but with a fault that the company financial crisis is a continuous process. This research used Multi-State Model to analyze company real financial status by comparing Backpropagation Neural Networks and Ordinal Logistic Regression Model. Besides the traditional financial indictors, We add the non-financial variables based on the theory of Cash Flow, Ownership Structure and CPA’s opinion to predict corporate financial distress. Then we use Factor Analysis or Stepwise Regression to choose the explained variables. To evaluate and discovery the stable model, We establish models every year with different methods during 2000-2004. In misclassification study, we use the type I error rate and Expected Cost of Misclassification (ECM) to evaluate the multi-state model.

This research established totally 45 models to predict financial distress. After evaluating by Accuracy, type I error and ECM, we find that the models built by Backpropagation Neural Networks is effective than Ordinal Logistic Regression Model in Multi-classification. And the models using Factor Analysis to choose explained indictors is highly stability with little error cost than using Stepwise Regression ones. Most of the variables chosen by Factor Analysis belong to Profitability, Capital Structure and CPA’s opinion. Otherwise Balanced sample training makes the Backpropagation Neural Networks more stable than unbalanced sample training. And the combined Network model using Fact analysis and balanced sample is even more much better than the others.

並列關鍵字

Ordinal Logistic Regression ； Financial distress prediction model ； Multi-classification ； Backpropagation Neural Networks

參考文獻

22. 黃振豐、呂紹強，「企業財務預警模式之研究-以財務及非財務因素建構」，2000，當代會計，第一卷第一期，19-40頁

1. Anurag Agarwal, Jefferson T.Davis, and Terry Ward,“Supporting Ordinal Four-State Classification Decisions Using Neural Networks,” Information Technology and Management, 2001, Vol.2, pp5-26.

3. Daniel E. O'Leary, “Using Neural Networks to Predict Corporate Failure,” International Journal of Intelligent Systems in Accounting, Finance & Management, 1998,Vol.7, pp187-197.

4. Harlan L. Etheridge, and Ram S. Sriram.,“A Comparison of the Relative Costs of Financial Distress Models : Artificial Neural Networks, Logit and Multivariate Discriminant Analysis,” International Journal of Intelligent Systems in Accounting, Finance & Management, 1997, Vol.6, pp235-248.

6. John Stephen Grice, Sr, “Reestimation of the Zmijewski and Ohlson Bankruptcy Prediction Models,” Advances in Accounting, 2003, Vol.20, pp77-93.