Please use this identifier to cite or link to this item:
|Title:||Temporal fuzzy cognitive maps for corporate credit ratings||Authors:||Zhong, Haoming||Keywords:||DRNTU::Engineering::Computer science and engineering||Issue Date:||2015||Source:||Zhong, H. (2015). Temporal fuzzy cognitive maps for corporate credit ratings. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||Corporate credit ratings are one of the key problems of the credit risk management, which has attracted much research attention since the credit crisis in 2007. Methodologies and models based on finance, statistics, and expert knowledge have been developed in past decades. The essential problem of corporate credit rating is finding out the relationships between financial factors and associating them to generate the credit assessment. Artificial Intelligence (AI) technologies such as neural networks and SVMs have demonstrated their remarkable performance on automatic credit ratings. However, the relationships between input financial factors and their rating results is not interpretable, which brings difficulties to justify and revise models. FCM is a hybrid approach combining advantages of expert systems and AI approaches, which makes it a promising approach for the credit rating problem. The work described by the thesis investigates three topics of FCM on the corporate credit rating problem, including (1) the FCM structure and the training algorithms, (2) relationships among FCM, correlation, and causation, and (3) temporalized FCMs. To study the capacity of FCM on credit rating, an experimental comparison study over the effectiveness of five lear~ing algorithms, i.e., BP, ELM, I-ELM, SVM, and FCM, is carried out targeting at comparing FCM with other AI approaches on the corporate credit rating problem. The effectiveness of the five algorithms is studied in terms of reliability and discrimination capacity. The experimental results show that Neural network based solutions, including BP, ELM, I-ELM, FCM, outperform SVM on reliability because they have better error distribution, while the SVM achieves a better performance on the discrimination testing. FCM, with a smaller structure than SLFN, obtains an equivalent performance as other approaches. It shows that the structure of FCM is more efficient than SLFN. To further study the capacity of FCM on credit rating with small sample size, an individual rating case is studied. Nokia, as a telecom corporation, has been downgraded from Al to Bl in past six years. Due to its wide rating variation in six years, Nokia is an appropriate example to test the performance of different AI models on single corporation analysis with incomplete sample data. By exploiting correlations between financial factors as priori knowledge, the FCM outperforms BP and SVM on the reliability. The result also indicates that SVM cannot identify cases which are not covered by the training samples. In addition, another experimental study has been carried out to compare the difference brought by the correlation combination. The study shows that the combination improves the reliability of the FCM and the consistency of FCM training results as well. The high consistency between trained FCMs facilitates the interpretation and justification of training results. Finally, the study demonstrates that there are strong temporal correlation between different financial factors. As one of the three causation rules, the temporal attributes play an essential role in the causality identification. FCM is a dynamic causal inferring tool. By temporalizing its relationships, FCM is able to support temporal causal inferring. The novel temporalized FCM with gradient descent learning algorithm is proposed to study the causal relationships between corporate financial factors. Three corporations, Nokia, Ericsson, and Google, are used to verify the proposed approach. The result shows that more than 50% of correlations have been eliminated by the training algorithm. The remained relationships, which are called "h-causations", are different from the initial correlations. An analysis on some "hcausations" indicates that they can be justified based on corporation financial conditions, which gives a subjective proof that "h-causations" are closer to causations than correlations. Based on the gradient descent algorithm, a temporalized FCM can be built by using machine learning. Meanwhile, FCM ca be built based on human expertise. A novel temporal FCM model, tFCM, is proposed to map human expertise to temporalized FCM with mapping patterns. An empirical study compares the difference between two kinds of causal relationships, the "if ... then ..." relationship and the "changes followed by changes" relationship, on error accumulations and sensitivities. Keywords: Corporate credit rating, ELM, I-ELM, SLFN, Neural network, SVM, FCM, Causality inference, Temporal fuzzy knowledge.||URI:||https://hdl.handle.net/10356/65676||DOI:||10.32657/10356/65676||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.