Using machine learning methods in financial distress prediction: An application for SMEs in Turkey

In this study, financial statements included between 2015-2018 of the SMEs consisted of 173 non-distressed and 219 distressed operating in Turkey have been examined . To predict to financial distress, it is intended to find the model that has the highest accurate prediction for each year before financial distress with using logistic regression, decision tree, random forest, support vector machines, K nearest neighbor and Naive Bayes model. Firms are considered distressed if bankruptcy or concordat (composition of debts) decision were issued by the competent courts in 2018. Empirical results indicate that for 1 and 2 years prior to financial distress, decision tree model is respectively the best classifier with overall accuracy of %90 and %97. Three years prior to financial distress, the Naive Bayes is the best classifier prediction model with overall accuracy of 97%. When the accuracy of all model results are examined, it has been obtained the higher successful prediction results when the further away from bankruptcy. Although it is the nearest term to distress year, to have lower model accuracy prediction rates in the t-1 year directed us to more detailed researches on the dynamics of the Turkey's economy in which operates firms. It has been observed that macroeconomic indicators (interest, inflation, exchange rate), acting similarly in 2015 and 2016, increased dramatically in 2017 and 2018. Distressed firms grew fast with high bank loans in 2017 and achieved high operating margin. However, these companies with low equity capital could not manage their financing expenses after a while. Distressed firms that could not make a profit with high bank debts and high financing expenses failed in 2018 because they could not meet their financial expenses. Distressed firms operate with high stock dependency ratio and low profitability. It has been determined that distressed firms have lower pre-tax profits, after-tax profits and return on equity than non-distress firms. In addition, it has been observed that the share of long-term liabilities in constant capital is higher. While the total foreign assets of non-distressed firms are twice their own funds, this situation is on average 4 times for distressed firms. It is observed that non-distressed firms operate with higher asset turnover and their operating volumes. In the period when cyclical fluctuations were high, it was observed that the models correct prediction rates were low. Therefore, we believe that adding more independent variables (internal and external factors affecting failure) to the model to achieve higher model accuracy rate will be beneficial in terms of obtaining more accurate and reliable results.