| Summary: | 碩士 === 國立臺灣科技大學 === 資訊管理系 === 106 === In late 2007, the subprime mortgage event in the United States caused the global housing markets and the stock markets to plummet. As a consequence, many companies lay off employees to reduce their operating cost. Ultimately, some companies could not withstand the crisis and distressed. Most of the existing researches considered only the financial soundness of a company in building a financial crisis prediction model. Recent researches considered corporate governance variables, macro-economic variables and external rating variables in building a prediction model.
Unlike the existing researches, in this thesis we considered the textual information in the daily news or posts on the webs in building the prediction model. We used text mining technique and convolutional neural network to extract from the daily news and web posts a new textual variable to build the prediction model. Furthermore, we collected totally 96 financial and non-financial prediction variables and then used the random forests method to select 15 most important variables as the final prediction variables. With the textual variable and the 15 selected variables, we built two different prediction models using Convolution Neural Network plus Support Vector Machine (CNN-SVM) and Convolution Neural Network plus Logistic Regression (CNN-LR). The experiment results showed that the CNN-SVM achieved the highest prediction accuracy of 92.51%. Furthermore, all the precision, recall and f-measure of the CNN-SVM were higher than 93%. We also compared our models with the other models which did not consider the textual variable. The experimental results showed that our models prevail over the other models in prediction accuracy.
|
|---|
