Prediction of company insolvency using data science methods

The main goal of the thesis was to create models designed to predict the insolvency of companies based on the maximum usability of available data. It means to use wide range of variables, depending on this, the different sizes of available data samples are derived, resulting in 10 experiments. The experiments were used for comparison from more interesting points of view, an example is comparing the predictive power of the model for different time periods, using different sets of variables and examining their importance, the influence of a data sample on model accuracy and changing the significance of variables, etc. Part of the thesis also compares the interpretability of the Random forest model and logistic regression. The work is based on data science methods, for this reason, the reader was briefly acquainted with this term and related terms in the introductory chapter. Furthermore, the CRISP-DM methodology was introduced, on which the research of the thesis was based. Some subsections of this methodology were theoretically analyzed in more detail – methods of reducing the number of variables or model selection and evaluation. The chapters contain mainly the theoretical basis of the methods used in this thesis. The thesis was also inspired by an article by Breiman (2001), based on which the models were selected according to two criteria - interpretability and predictive power. The model of penalized logistic regression was used as a representative of the interpretable model, and the random forest model was chosen for its predictive power. A number of specific techniques have been used to improve the accuracy of the model, such as stratified cross-validation, the Oversampling method, or Grid Search parameter tuning. Text mining methods were also used in the work. Data processing took place in Python, modeling in R.