No abstract

No abstract

No abstract

Since the overdue amount of credit cards has been increasing year by year, the rising credit card delinquencies might prevent the commercial banks to allocate more funds in profitable investments. At the same time, the high processing costs of credit card delinquencies through manual verification also affect the competitiveness of credit card issuers. Because the market competition becomes strict in the era of financial technology, to predict correctly whether cardholders will be unable to pay off credit card debt and establish an effective risk prediction model is the major purpose of this study.We first implemented four machine-learning approaches to predict the default cases, however, most models encountered challenges to resolve imbalance problem of delinquency cases in data sets and reported lower predictive accuracy. Two inference strategies including grey incidence analysis and fuzzy decision tree were proposed to improve the predictive performance. The average accuracy for both strategies were increased from 0.82 to 0.86 and 0.89 respectively. In addition, the deep learning approach integrated with various network structures was also incorporated to evaluate model performance. The experiment results indicated the deep neural network performed better in most evaluation metrics and achieved an impressively high accuracy of 0.93 if compare to the machine learning models. Finally, three feature selection methods were employed with the deep learning model, and the results showed similar predictive accuracy as the original deep learning models with slightly better performance being reported by filtering variables with the grey incidence analysis. This research work could be extended to apply more complicated deep learning algorithms to learn and trace the behaviors of the credit card holders and reduce the default risks for banking industries.Index Terms-Default prediction, machine learning, deep neural network, deep learning.

Value investing is one of the most popular investment strategies for investors to search for the undervalued stocks based on their financial reports and balance sheets. However, the numerous metrics derived from the financial statements are not easy for the investor to analyze and determine the financial health of a company. The main purpose of this study is to employ feature extraction to identify a smaller number of financial ratios for the prediction of stock return which reflects the quality of a company. Two regression approaches, including Multilayer Perceptron model and Meta Regression by discretization model, were incorporated with feature extraction to evaluate the forecast performance for two different industries in Taiwan. The results demonstrated that the prediction errors were improved for both models by the feature extraction strategy which reduced the original 16 variables to5 variables. Besides that, both models achieved better prediction result in concrete industry rather than rubber industry. Finally, this paper concluded that the overall performance of the Multilayer Perceptron model is better than the other model.