Assessment of consumer credit risk via random forests method improved with a combined meta-heuristic approach

Global financial crisis, which dates back to 2007-2009, raised a growing interest to credit risk assessment analyses. Long-established and big banks became a source of systemic risk and caused their exposure to debt risks to be transmitted throughout the whole financial system by landing mortgages to too many households with low repayment capabilities, which are also known as subprime mortgages. History of modern studies for collecting information and assessment about credit applicants have begun around 1940s but first examples of statistical methods usage showed up in late 1960s. From that day on, in parallel to rapidly increasing amount of credit and credit borrowers, much more data has been collected. Along with progress made in computer science, advanced methods which are based on intensive computation to analyze these data are also developed. Needs and solutions are mutually interacted in a spiral way up to the present and a wide literature has been emerged about credit applications' assessment. Most of the literature is consisted of statistical and machine learning methods. Most of these methods are very successful for deriving interpretable information and correct predictions from complex datasets with many variables and observations. However, little differences will have great impacts on company profits or losses and even on the sustainability of the system. That's the reason for continuing efforts to improve and develop existing methods and approaches. As a part of these efforts, in this study, a well-known consumer credit dataset which is used as a benchmark set for comparison purposes, is analyzed with a hybrid approach. Dataset learning and predictions are conducted with a machine learning method called Random Forests, which has been used intensely in subjects like biology and medicine but became used in finance recently. In order to improve the method, feature selection is applied with Genetic Algorithm and Simulated Annealing meta-heuristic approaches. Aim is to eliminate irrelevant variables that disturb learning process and cause worse predictions. Random Forests method is embedded to the design of combined meta-heuristics proposed by this thesis. Combined design allowed to use strengths of meta-heuristics while overcoming weaknesses. In credit assessment studies, Genetic Algorithm is widely used but there is no example of a study with Simulated Annealing. Results show that hybridized implementation of Random Forests produces statistically significant increases in prediction performances vis a vis sole usage. When compared with other studies adopted the same dataset, it is seen that proposed hybrid method shows high performance. Method will be highly beneficial for decision makers in assessment of credit applications.