近年全球經濟不景氣，不少企業陷入財務危機的考驗之中，部分企業為了存續，轉而向銀行貸款，然而，如果這些企業的償債能力不足，他們仍可能面臨破產，更甚者將無法償還向銀行借貸的現金。為了避免此行況，本篇論文旨在透過財務變數以及自定義的「輿情風險分數」，來預測企業的信用評等與財務危機。文字探勘上，使用了Latent Dirichlet allocation主題模型演算法，從風險新聞當中萃取出六大分類的「風險詞庫」，並透過「風險詞庫」計算出「輿情風險分數」；接著透過深度類神經網路，建構「TCRI信用評等預測模型」以及「財務危機預測模型」等兩個深度學習模型。
在TCRI信用評等預測模型，本論文比較了添加文字探勘特徵值「輿情風險分數」前後的模型正確率，添加輿情風險分數之後的正確率比起尚未添加前，高出至少4個百分點，正確率高達84%；在財務危機預測模型方面，本論文將TCRI信用評等作為重要的財務危機因子，比較添加前與添加後的模型正確率，在添加TCRI信用評等之後，模型正確率提升超過4個百分點、高達97%。
研究結果顯示，文字探勘所產生的特徵值有助於從文字資料中，找出一家企業潛藏的風險，而作為頗具公信力的信用評等指標，TCRI對於財務危機的預測確實有顯著的幫助。

Due to the global economic downturn, many companies have faced financial distress, leading to the crisis of bankruptcy. Some of the distress companies will lend to banks in order to maintain sufficient cash flow. However, if they have insufficient solvency, they may still face bankruptcy, which will make it impossible for the company to repay the bank.
To avoid this situation, this paper aims at predicting companies’ financial distress on several critical financial variables and a self-defined feature: risk score, a sentiment feature that uses Latent Dirichlet allocation (LDA) topic modeling to determine the potential risks implied in news articles. We first use text mining algorithm to preprocess our sentiment data, select important keywords by LDA and generate 6 risk corpus. With financial and news data, we then construct 2 prediction model. One is TCRI (Taiwan Corporate Credit Risk Index) prediction model, where TCRI is an iconic index that over 90% of banks refer to TCRI for credit lending in Taiwan. The other is the financial distress prediction model. As a crucial factor, TCRI will be one of the features of the financial distress prediction model.
Both these 2 models use deep neural network. We conduct several times to find the best parameter settings. In the first TCRI prediction model, we compare 2 versions of prediction mode: TCRI with risk score and TCRI without risk score. The former model achieves over 84% in accuracy and higher than the latter model. In the second financial distress prediction model, we compare 2 versions of prediction mode: financial distress prediction with TCRI and without TCRI and 4% accuracy more than the case without TCRI. The result shows that TCRI plays an important role with over 97% accuracy in average.
We conclude that text data such as news article can help disclose the potential signs of a company’s crisis, getting a better result. Also, a credible indicator, such as TCRI, can improve the accuracy of financial distress prediction.

[1] H. Beaver, W. (2010). Financial Statement Analysis and the Prediction of Financial Distress (Vol. 5).
[2] 台灣經濟新報. TEJ TCRI 台灣企業信用風險指標. from https://www.tej.com.tw/tcri/
[3] Beaver, W. H. (1966). Financial Ratios As Predictors of Failure. Journal of Accounting Research, 4, 71-111. doi: 10.2307/2490171
[4] Altman, E. I. (1968). FINANCIAL RATIOS, DISCRIMINANT ANALYSIS AND PREDICTION OF CORPORATE BANKRUPTCY. Journal of Finance, 23(4), 589-609. doi: 10.2307/2978933
[5] Deakin, E. B. (1972). DISCRIMINANT ANALYSIS OF PREDICTORS OF BUSINESS FAILURE. Journal of Accounting Research, 10(1), 167-179. doi: 10.2307/2490225
[6] Blum, M. (1974). FAILING COMPANY DISCRIMINANT-ANALYSIS. Journal of Accounting Research, 12(1), 1-25. doi: 10.2307/2490525
[7] Ohlson, J. A. (1980). Financial Ratios and the Probabilistic Prediction of Bankruptcy. Journal of Accounting Research, 18(1), 109-131. doi: 10.2307/2490395
[8] Zmijewski, M. E. (1984). Methodological Issues Related to the Estimation of Financial Distress Prediction Models. Journal of Accounting Research, 22, 59-82. doi: 10.2307/2490859
[9] Huang, T.-H., Leu, Y., & Pan, W.-T. (2016). Constructing ZSCORE-based financial crisis warning models using fruit fly optimization algorithm and general regression neural network. Kybernetes, 45(4), 650-665. doi: doi:10.1108/K-08-2015-0208
[10] Sullivan, D. (2001). Document Warehousing and Text Mining: Techniques for Improving Business Operations, Marketing, and Sales: John Wiley & Sons, Inc.
[11] Gupta, V., & Lehal, G. (2009). A Survey of Text Mining Techniques and Applications (Vol. 1).
[12] Blei, D. M. (2012). Probabilistic topic models. Commun. ACM, 55(4), 77-84. doi: 10.1145/2133806.2133826
[13] Moro, S., Cortez, P., & Rita, P. (2015). Business intelligence in banking: A literature analysis from 2002 to 2013 using text mining and latent Dirichlet allocation. Expert Systems with Applications, 42(3), 1314-1324. doi: https://doi.org/10.1016/j.eswa.2014.09.024
[14] Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet allocation. Journal of Machine Learning Research, 3(4-5), 993-1022. doi: 10.1162/jmlr.2003.3.4-5.993
[15] Ganegedara, T. (2018). Intuitive Guide to Latent Dirichlet Allocation. Retrieved 2019/5/12, from https://towardsdatascience.com/light-on-math-machine-learning-intuitive-guide-to-latent-dirichlet-allocation-437c81220158
[16] 陳惠玲、李冠皓. (2009). TCRI製作程序. TEJ信用風險評估專刊.
[17] Rosenblatt, F. (1958). The perceptron: a probabilistic model for information storage and organization in the brain. Psychological review, 65(6), 386.
[18] Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323(6088), 533-536. doi: 10.1038/323533a0
[19] Grün, B., & Hornik, K. (2011). topicmodels: An R Package for Fitting Topic Models. 2011, 40(13), 30. doi: 10.18637/jss.v040.i13
[20] Han, J., Kamber, M., & Pei, J. (2012). 8 - Classification: Basic Concepts. In J. Han, M. Kamber, & J. Pei (Eds.), Data Mining (Third Edition) (pp. 327-391). Boston: Morgan Kaufmann.
[21] 曾冠人. (2004). 應用一般迴歸神經網路法構建財務危機預警模式. (碩士), 國立交通大學, 新竹市. Retrieved from https://hdl.handle.net/11296/9d6wtx
[22] 粘縝. (2011). 應用資料採礦與果蠅演算法建構財務危機預警模型－以台灣上市櫃公司為例. (碩士), 東吳大學, 台北市. Retrieved from https://hdl.handle.net/11296/k6689n
[23] 李振瑋. (2018). 結合卷積神經網路與支援向量機之企業財務危機預測模型. (碩士), 國立臺灣科技大學, 台北市. Retrieved from https://hdl.handle.net/11296/46ga2v
[24] Stewart, M. (2019). The Actual Difference Between Statistics and Machine Learning. Retrieved 2019/5/2, from https://towardsdatascience.com/the-actual-difference-between-statistics-and-machine-learning-64b49f07ea3