Bankruptcy prediction using genetic programming - a case study of Hungarian accommodation provider firms

Absztrakt (kivonat)

Bankruptcy affects every member of the society and the economy. Bankruptcy alters the structure of the economy, it is costly and in general it has a negative connotation. All the same, bankruptcy is useful if viewed from a broader perspective, as firms incapable of living have no place among the healthy enterprises. This necessity does not mean that the phenomenon is not worth investigating. Bankruptcy prediction offers an opportunity to economic actors to make their decisions based on facts. On the one hand it is of economic interest of the decision maker, on the other hand such analyses can be enforced by regulators as well. The goal of my thesis is to introduce a modeling method along a real world example which utilizes a not widespread approach. It is my purpose to do this by placing the method in context, introduce it in general, build a model with it and compare its performance with a well-known and widely accepted alternative. My thesis is divided into three main parts. The theoretical part aims to reveal the most important early warning methods available. A historical review reflects the progress in this field and introduces terms that are essential to understand bankruptcy prediction. The theoretical part begins with the classical methods that build mostly on statistics, the second covers methods that use computational intelligence, while the third separately introduces the achievements of Hungarian authors. Genetic programming belongs to the second subchapter, to evolutionary algorithms. By illustrating this family with examples I highlight the areas of life and science where such methods have excelled. In the methodological part I introduce the binary logistic regression which will serve as the reference for comparison. In this part I also go into detail about genetic programming and describe methods that are used for model performance comparison later on. As part of the empirical research I make an overview about the status of Hungarian accommodation provider firms and define the path that firms may take towards non-existence in Hungary. I do not make a distinction between the different legal terms and treat all of them equal. These preliminary steps are necessary to formulate my hypotheses precisely. My first hypothesis investigates whether a genetic programming model can be an effective bankruptcy predictor, while the second hypothesis questions whether such a model can outperform a binary logistic regression. I illustrate the process of model building in detail. The input data of the models are Hungarian accommodation provider firms. I gather publicly available data, calculate financial ratios and prepare them for modeling. Due to the low number of observations I create more samples for model building and make my comparison based on them. I not only summarize the results of the models in plots but also reveal the process by noting the exact steps on a random sample. In case of the logistic regression I use SPSS while in case of genetic programming I discuss my own algorithm written in python. Having compared the results of the models I have arrived at the conclusion that genetic programming is capable of bankruptcy prediction, what is more it can outperform a logistic regression. As a result I have accepted both my hypotheses. However, I would like to highlight that my results are based on a small sample and thus cannot be used to draw general conclusions about performance. In conclusion I have found that genetic programming has proven its reason for existence in bankruptcy prediction even on Hungarian data. It provides both academics and practitioners of this field with many research questions. I have highlighted further research directions along with possible causes that pose difficulties in spreading the method and using it in practical applications.