考慮總體經濟與產業因素之違約預測模型—以台灣半導體產業為例

Abstract

針對倒帳機率的預測，除了會計變數於學術界與實務界大行其道外，許多研究皆指出總體與產業因素對於倒帳機率亦具有預測能力；然而，卻鮮少有研究針對美國以外地區個別產業倒帳機率的預測因子進行探討分析。本研究目的在於找出預測台灣半導體產業倒帳機率的最佳總體與產業變數，並建立該產業之倒帳預測模型。 本研究根據需求、供給、競爭態勢與總體經濟等四項層面分別挑選變數。研究樣本包括2001年1月至2006年12月之84筆月資料，涵蓋於台灣註冊之積體電路製造商與積體電路設計業者。研究首先分析自變數與應變數之間的關係，並進行簡單迴歸分析，接著將具顯著解釋力之變數納入複迴歸分析，以建立三個月期與一年期之預測模型。結果顯示，數項總體與產業變數對於半導體產業違約機率具有顯著解釋力，且三個月期與一年期違約預測之最佳模型並不相同。整體而言，本研究結論如下：一、在本研究所有情況下，半導體設備訂單額與OECD領先指標變動值對於倒帳機率皆具有顯著解釋力；二、在一年期的預測模型中，數項變數與違約機率具有非線性關係，顯示其對於倒帳之影響因程度不同而有所改變；三、在針對積體電路製造商所建立之模型與針對積體電路設計業者所建立之模型中，兩者所包含的變數有所不同，即便是同一變數，與倒帳機率之間的關係也存在著差異；四、全產業模型在三個月期違約預測中誤差較小，而針對次產業分別建立之模型則在一年期預測中表現較佳。

Indicative roles of macroeconomic and industrial attributes in default probability prediction have been reported by many studies. However, studies conducted to identify the best indicators for individual industries outside US are scant. Therefore, the purpose of this paper is to identify macroeconomic and industrial variables and to build default prediction models for Taiwan’s semiconductor firms. This study selects variables through demand, supply, competitive conditions, and macroeconomic aspect. The samples included are 84 sets of monthly observations over the period from January 2000 to December 2006 covering Taiwan registered IC manufacturers and IC design houses. After analyzing the relationship between dependent and independent variables, the significant variables identified in simple regressions are included to build models for three-month and one-year predictions. Results of this study showed that several variables have significant explanatory power and the models which perform better in three-month and one-year prediction are different. To conclude, four findings may add to our understanding of default prediction for Taiwan’s semiconductor firms. (1) The percentage change in the semiconductor equipment booking value and the change in composite index of leading indicators of OECD appear to be the most pronounced variables to all default probabilities. (2) Dummy slope coefficients for several variables are significant in the one-year default prediction models, implying that the effects of these variables change as their level increase. (3) Several variables included in models for IC manufacturers only and IC design houses only are different and diverse relations with the same indicators exist. (4) The model for the whole industry performs better in predicting defaults in three-month prediction, while the models for separated groups should be employed in one-year prediction.