| Abstract [eng] |
The aim of this master‘s thesis is to analyse corporate bankruptcy prediction as one of the key challenges in business and finance. Business failure can lead to direct and indirect losses for various stakeholders, including investors, creditors, corporate managers, and regulatory authorities. Traditional bankruptcy models, that mostly rely on financial ratios can often lose accuracy in a constantly changing economic environment. Therefore, it is relevant to develop dynamic and localised models capable of integrated both firm-level financial data and national macroeconimic factors, reflecting business cycle phases related to systemic risk. The object of this research is potential ways of predicting bankruptcy risk of Lithuanian companies. The main goal of this work is to analyse research on corporate bankruptcy prediction and, based on this analysis, develop the most accurate bankruptcy prediction model. This model for Lithuanian company data has to combine both, financial and macroeconomic indicators. In order to achieve this goal, the following tasks were set: to analyse the theoretical aspects of corporate bankruptcy prediction, identifying the factors that determine bankruptcy risk, the characteristics of the models used, and the impact of data structure and the macroeconomic environment on the accuracy of predictions; to compile a final dataset of Lithuanian companies; perform detailed data preparation, ensuring data integrity and handling missing values; to analyse the applicability of logistic regression, random forests and gradient boosting algorithms, and evaluate the accuracy of the developed models using standard metrics; to select the most accurate prediction model, and provide practical recommendations, according to the given results. The methodology of this work is based on both traditional statistical and advanced machine learning methods. Real data of Lithuanian companies from the period 2013-2022 was used for the study. The final dataset consists of 1454 bankrupt and 39011 non-bankrupt companies. In order to objectively evaluate the reliability of all the models, the data was split chronologically into two periods. 2013-2020 period (containing 1024 bankruptcy cases) for training the models and 2021-2022 period (containing 430 bankruptcy cases) was used for testing. The final data used for the models consists of 81 independent financiel variables (covering profitability, liquidity, solvency, operational efficiency) and key macroeconomic indicators (GDP growth, EURIBOR, consumer price index and unemployment rate) with lags and changes. During the data preparation phase, due to a severe class imbalance in the data, MICE method was applied. Classification problem was solved by using and comparing several models. Logistic regression (with and without elastic net regularisation), random forests and GBM. For the evaluation, the standard metrics were used: AUC, sensitivity (recall), F1-score. The results of the study demonstrated that logistic regression (a traditional statistical method) is ineffective when dealing with severe class imbalance and it fails to accurately identify bankrupt companies. Meanwhile, machine learning algorithms significantly outperform traditional models. The deep GBM model demonstrated the best performance, achieving a high AUC score of 0.9104 and highest sensitivity (recall) of 0.9. This indicates that this model is capable of succesfully identifying 90 % of companies that are about to bankrupt within a one-year horizon. Analysis of variable importance revealed that financial reporting delay (reporting_delay) is a critical internal risk signal for a company. Furthermore, indicators related to equity, assets and liabilities also have a significant impact on bankruptcy prediction. The integration of macroeconomic factors provided models with an analytical advantage in assessing systemic risk. During the analysed period, the methods successfully identified specific risks related to business cycle factors: labor market overheating (unemployment rate changes) and rising raw material prices. The results of the study prove that bankruptcy prediction cannot be separated from the macroeconomic context. The combination of micro and macroeconomic variables ensures a higher prediction accuracy. |
