考量市場與會計基礎模型於營建產業財務危機預測之研究

Abstract

營建產業相對於其他產業，擁有較高的財務風險，由於營造產業之承攬業務金額龐大且工期長，對國內外的政府機構、業主、貸款銀行、保險公司和承包商來說，評估營建產業公司的財務危機機率是一個非常重要的課題。早期發展之財務危機預測模型，多是針對所有產業，鮮少針對個別產業進行研究。主要的原因是，集中於單一的行業的研究樣本蒐集困難，尤其是危機公司的樣本數與財務正常樣本數相較下過於稀少，而營建產業由於其獨特的財務特性更是大多被排除在早期的研究範圍外。有鑑於營建產業的產業特性及編列財務報表之會計處理原則與其他產業不盡相同，目前的營建產業環境也急需針對營建產業財務危機預測之研究，故本研究目的即為建立營建公司財務危機預測模型。 傳統之財務危機預測模型以會計基礎模型為主，學者們假設在財務危機公司與財務正常公司的會計財務報表應該有所不同，並試圖利用一些資料探勘或迴歸方法來找出這些不同，因此會計基礎模型需要大量的歷史樣本做為參考。過去有關會計基礎模型財務危機預測的研究中，建立樣本集時經常採用配對法，即一個財務危機樣本配對一至兩個財務正常樣本。由於現實中財務危機為相對稀少的事件，如此建立樣本集的方式會造成選樣偏誤，近來已有學者提出應將所有可得之樣本皆放入樣本集內。如此一來便帶來了新的問題：財務正常樣本的數量遠多於財務危機樣本，這樣的不平衡被稱為「分類間不平衡」。會計基礎模型使用的迴歸方法常常只能表現樣本集中佔大部分的樣本特性，忽略佔小部份的樣本。在財務危機預測中，佔小部分的財務危機樣本卻才是估計準確的關鍵，因此在含有分類間不平衡的樣本集中，會計基礎模型的預測能力就受到了限制。為了改善這個分類間不平衡的問題，本研究目的其一即為應用兩項重覆取樣技術：「強化訓練」及「Synthetic Minority Over-sampling Technique (SMOTE)」以改善此問題，其目的在於增加財務危機樣本的數量，減少分類間的不平衡。 除了會計資訊外，近年來也有學者開始使用以公司股價為主要資訊來源的市場基礎模型。在效率的股票市場中，股價應能充份表現公司價值，是另一個優質的財務資訊來源。本研究目的其二即為建立一整合會計資訊及股票市場資訊之混合型模型，並搭配重覆取樣技術以提高預測模型的準確度。 本研究採用實證的方式評估模模型的預測能力，收集了美國與台灣營建產業公司財務樣本，分別建立美國及台灣的營建產業財務危機預測模型，以了解不同市場下對模型的影響。由實證的結果可知，比起單獨使用會計基礎模型或市場基礎模型，混合型模型有更高的預測能力。再搭配重覆取樣技術之後，還能進一步提昇會計基礎模型及混合型模型的預測能力。藉由這些技術的應用，能及早預測營建產業可能發生之財務危機公司，以提供營建產業經營者、管理者、金融機構、保險公司，投資大眾及營建相關產業等做為參考，更具體地瞭解及正確地辨別營建公司之財務危機。

Due to the special financial characteristic of construction industry, past researches on bankruptcy prediction models mostly excluded the construction industry from their sample. However, the financial health of construction contractors is critical in successfully completing a project. The financial default probability of the construction industry is always an important issue for governmental organizations, construction owners, lending institutions, surety underwriters, and contractors. Thus, this research aims to measure and predict the construction contractor default risk. The financial default predicting models developed in past literatures are in large built by historical accounting information. They were called as “accounting-based models”. These researches supposed that there may be different patterns between defaulters and non-defaulters in historical accounting information, and tried to find out these patterns by some regression or data mining analysis. Thus, scholars usually need numerous of samples to build accounting-based models. Most of the previous studies on prediction construction contractor default used sample-match method to build their sample set, which produces sample selection biases. In order to avoid the sample selection biases, this research used all available firm-years samples during the sample period. Yet this brings a new challenge: the number of non-defaulted samples greatly exceeds the defaulted samples, which is referred to as between-class imbalance. Accounting-based models only demonstrate the distribution of the major parts of input points, ignoring the small parts of input points. Thus using the accounting-based models on default prediction with imbalance data set is not satisfactory. The primary objective of this research is to improve this shortcoming by 2 kinds of over-sampling technique: “replication” and “Synthetic Minority Over-sampling Technique (SMOTE)”. The purpose of these over-sampling techniques is to increase the number of default samples, and reduce the between-class imbalance. Besides the accounting-based models, the option-based model is another way to predict company default. The option-based model doesn’t catch the information by data mining, but depicts the physical mechanism of company’s default by using option-pricing equations with the main input: company stock price. In an efficient market, the company’s stock price could be a good source of information because it not only reflects accounting and economic information but also reflects qualitative factors such as management and technique. The second objective of this research is to build hybrid models which combine accounting and stock market information. The empirical results of this research show that the hybrid models outperform the accounting-based models and the option-based model. With the over-sampling techniques, the predicting performance of models could be even better. Thus, this research recommends the proposed hybrid models with over-sampling techniques as an alternative to the traditionally used models.