我國銀行金融預警模型之實證研究

Empirical Study on Financial Early-Warning System Models for Taiwan Banks

李明萱(Ming-Shiuan Lee)

指導教授 ： 林嬋娟

國立臺灣大學/管理學院/會計學研究所/碩士(2005年)

https://doi.org/10.6342/NTU.2005.02959

摘要

近年來政府積極研修金融法規並強化金融監理制度，並將民國90年度訂為「金融改革年」加速推動多項重大金改措施，銀行績效評量與監理日益重要。本研究期建立一金融預警模型，採用逐步Logistic迴歸之混合逐步選擇法，並利用Logistic累積機率函數計算各銀行落入經營體質不佳群組之機率，該模型之應用除可作為現行「申報資料排序系統」監督銀行經營的輔助工具、金融監理單位分配金檢資源之參考外，所推估各銀行經營不佳之機率值亦可作為存款保險風險差別費率計價之依據。
本研究採用兩種方式定義問題銀行，分別建立模型A與B，再細分為八期間模型。模型A：參考West (1985) 以及我國「存款保險差別費率實施方案修正案」，以「檢查資料評等綜合得分」定義「經營體質不佳」群組，建立單年度模型88-A、89-A、90-A及三年度混合模型88-89-90-A；模型B：從我國類似美國立即導正措施規定中選取資本適足率等三種主要之銀行監理財務指標，定義「經營體質不佳」群組，建立單年度模型88-B、89-B、90-B及三年度混合模型88-89-90-B。
經本研究測試，模型A及B皆具有良好正確分類能力，其中單年度模型又較三年度混合模型為佳，正確分類率介於89.13%至95.56%間，加權效率性指標亦平均達66.65%，與以往文獻相較乃具水準之上。就模型預測能力而言，模型A較模型B為佳，且單年度模型相較於三年度混合模型為佳，正確預測率達73.91%至82.61%間，加權效率性指標可達53.40%。實證結果亦發現：(1) 樣本涵蓋的期間不宜過長，否則模型無法準確反映實況，建議以單年度資料建立之模型將優於以數年度混合資料所建立之模型；(2) 模型90-A較模型88-89-90-A與預測年度91年時間相近，故具較佳的預測能力，此意味預警模型之建構是一項持續性工作，相關指標與權數等之採用應隨整體金融環境與經營風險之變遷，予以檢視並進行調整；(3) 美國CAMELS評等系統所涵蓋之各評量構面皆攸關銀行體質良窳不可偏廢；(4) 任何統計預警模型極少能達百分之百之正確率，故在使用上宜注意其各年度之趨勢變化與分析，而非僅著重其單一年度結果，同時應併同其他資訊交互驗證，以提升模型之實用性。

關鍵字

金融預警模型 ； 銀行業 ； Logit model ； Logistic model ； CAMELS評等系統

並列摘要

Recently Taiwan’s government has positively enacted financial bills and strengthened banking supervision while labeling 2001 as “the Year of Financial Reform” to expeditiously promote several critical measures of financial improvement. As the assessment and supervision of bank’s performance are getting significant, this study intends to build a stepwise logistic regression model for the Financial Early-Warning System (hereinafter referred to as EWS model) and takes advantage of its cumulative probability function to estimate the probability of being classified as the group with unsatisfactory condition for each bank. This model can serve as a possible guide to bank monitoring in an auxiliary role for the current “Call Report Percentile Ranking System” as well as can be a reference for banking regulators to efficiently allocate examination resources. Moreover, in terms of “Deposit Insurance Risk-based Premium System”, the risk level of individual insured institution can be determined by the probability derived from such model.
In the study, model A and model B are respectively established by different definitions of problem banks. According to West (1985) and Taiwan’s “Implementation Scheme for the Deposit Insurance Risk-based Premium System”, it is single-year model 88-A, 89-A, 90-A, and three-year model 88-89-90-A that are built on the bases of “Composite scores of the Examination Data Rating System”. In contrast, to build single-year model 88-B, 89-B, 90-B, and three-year model 88-89-90-B, three main financial indicators, such as “capital adequacy ratio”, etc., are chosen from those in Taiwan’s prompt corrective action to identify problem banks.
The empirical results show that both model A and B, especially the single-year models, are able to correctly identify the problem banks of the original samples with higher percentage of correct classifications, which is from 89.13% to 95.56%, and with better average weighted efficiency indicator, which is around 66.65%. With regard to the prediction results using holdout samples of next year, model 88-A, 89-A, and 90-A outperform model 88-89-90-A and model B in every test in terms of higher percentage of correct classifications and weighted efficiency indicator, which are from 73.91% to 82.61% and up to 53.40% respectively. Moreover, the study finds that (1) in order to increase the prediction accuracy of EWS models, the data set chosen from one-year period to build the models is better than that from three-year period. (2) Timely modifications and revisions of EWS models are necessary to meet with the rapid changes in the financial market and operational risks. (3) All the areas of CAMELS rating system of the US Federal Financial Supervisory agencies are relevant to the assessment of banks’ operational condition. No one can be neglected. (4) Due to the inherited limitation of predictive errors, it is appropriate to analyze and emphasize the long-term trend of models’ results rather than just to focus on short-term results, which might happen to be an error. The EWS model must also be complemented by on-site examination and other methods to enhance its usefulness.

並列關鍵字

Financial Early-Warning System models ； banks ； Logit model ； Logistic model ； CAMELS rating system

參考文獻

35. Altman, E. I., 1968, Financial ratios, discriminant analysis and the prediction of corporate bankruptcy, The Journal of Finance, vol.23, no.4, Sep., 589-609.

36. —————, R. G. Haldeman, and P. Narayanan, 1977, Zeta analysis：A new model to identify bankruptcy risk of corporations, Journal of Banking & Finance, 1, 29-54.

44. Espahbodi, P., 1991, Identification of problem banks and binary choice models, Journal of Banking & Finance, Vol.15.No.1,pp.53-71.

45. Gessner, G., W. A. Kamakura, N. K. Malhotra, and M. E. Zmijewski, 1988, Estimating models with binary dependent variable: Some theoretical and empirical observations. Journal of Business Research, 16(1)：49-65.

47. Hwang, D. Y., C. F. Lee, and K. T. Liaw, 1997, Forecasting bank failures and deposit insurance premium, International Review of Economics and Finance, 6(3): 317-334.