本研究延續Shumway (2001) 從存活分析(Survival Analysis)觀點切入，利用離散型風險模型(Discrete-time Hazard Model)──亦即Shumway 所稱之多期邏輯斯迴歸模型(Multi-period Logistic Regression Model)，建立企業財務危機預警模型。研究選取1986 年至2008 年間718 家上市公司，其中110 家發生財務危機事件，共計6,782 公司/年資料 (firm-year)。有別於Shumway 提出的Log 基期風險型式，本文根據事件發生率圖提出Quadratic 基期風險型式，接著利用4組（或基於會計測量，或基於市場測量）時間相依共變量 (Time-dependent Covariate)建立2 組離散型風險模型(Log 與Quadratic)，並與傳統僅考量單期資料的邏輯斯迴歸模型比較。實證結果顯示，離散型風險模型的解釋變數與破產機率皆符合預期關係，而傳統邏輯斯迴歸模型則有時會出現不符合預期關係的情況；研究亦顯示離散型風險模型預測能力絕大多數情況下優於傳統邏輯斯迴歸模型，在所有模型組合中，以Quadratic 基期風險型式搭配財務變數、市場變數的解釋變數組合而成的離散型風險模型，擁有最佳預測能力。 / Based on the viewpoint of survival analysis from Shumway (2001), the presentthesis utilizes discrete-time hazard model, also called multi-period logistic regression model, to forecast corporate financial distress. From 1986 to 2008, this research chooses 718 listed companies within, which includes 110 failures, as the subjects, summing to 6,782 firm-year data. Being different from Shumway’s log baseline hazard form,we proposed to use quadratic baseline hazard form according to empirical evidence. Then, four groups of time-dependent covariates, which are accounting-based measure or market-based measure, are applied to build two sets of discrete-time hazard model, which is compared with the single-period logistic regression model. The results show that there exists the expected relationship between covariates and predict probability in discrete-time hazard model, while there sometimes lacks it in single-period logistic regression model. The results also show that discrete-time hazard model has better predictive capability than single-period logistic regression model. The model, which combines quadratic baseline hazard form with market and accounting variables, has the best predictive capability among all models.

本財務預警模型研究延續Shumway(2001)年所提出的離散型風險模型(Discrete-time Hazard Model)架構，即Shumway 所稱之多期邏輯斯迴歸模型(Multiperiod logistic regression model) ，來建立銀行財務預警模型。不同於Shumway所提出的Log 基期風險式，研究者根據實際財務危機發生機率圖提出Quadratic 基期風險式。由於離散型風險模型考量與時間相依共變量(Time-dependent covariate)，該模型可以納入隨時間變動的的市場與總體變數，這是單期模型無法達到的。實證結果顯示，不論是否有加入總體與市場變數，Quadratic 基期風險式離散型模型在樣本內檢測表現都比單期模型與Log 基期風險式離散型模型好，研究亦顯示樣本外的預測Quadratic基期風險式在大多數情況都優於Log 基期風險式與單期模型 / This paper continues Shumway(2001) studies on discrete time hazard model, the so called multi-period logistic regression model, to develop a bank failure early warning model . Different from log baseline hazard form proposed by Shumway, author present quadratic baseline hazard form based on the pattern of real default rate. By incorporating time-varying covariates, our model enables us to utilize macroeconomic and market variables, which cannot be incorporated into in a one-period model. We find that our model significantly outperforms the single period logit model and Log baseline hazard model with and without the macroeconomic and market variables at in-sample estimation. The improvement in accuracy comes both from the time-series bank-specific variables and from the time-series macroeconomic variables. Our research also shows that quadratic baseline hazard model outperforms Log baseline hazard model and single period logit model in out-of-sample prediction.