Вankruptcy Forecasting Based on Textual Information

The problem of bankruptcy and how it can be predicted in advance seems to be crucial not only for company's owners, but for investors, analysts and groups of stakeholders as well. While the quantitative models built on financial values seem to occupy almost the whole niche for financial failure forecasting, they cannot be considered to be the only effective tools for predicting as forecasting based on qualitative information is becoming more and more popular. The main aim of the study is to analyze the annual reports of retail companies from various countries in order to identify, whether the bankruptcy can be forecasted by using the text-based metrics. The metrics included texts’ vectors, informational entropy and emotional negativity and positivity, which were measured by using Harvard IV-4 Dictionary and Henry’s financial list of emotional words. Two sets of models produced on dictionaries were compared between each other in order to choose the list which showed the higher accuracy. On the basis of the calculated metrics, several machine learning models were built. The performance of each model it terms of accuracy, precision and recall was evaluated, and the comparison with the quantitative method, Altman's Z-score model, mas made. It was found out that the combination of mentioned metrics with such algorithms of machine learning as Support Vector Machines, K-Nearest Neighbors, Random Forest and Gradient Boosting provides higher accuracy of prediction than widely used Altman’s Z-score model. Besides from that, the models with sentiments calculated via Henry’s dictionary of financial vocabulary demonstrated better performance than the algorithms including sentiments from Harvard IV-4 dictionary. Overall, 170 reports of bankrupted and non-bankrupted firms within the period from 2010 to 2019 were collected and processed to make the sample balanced and reduce the impact of the world financial crisis. This project focused on industry specifics instead of more widespread approach of national context analysis. This may allow future researchers to explore and compare bankruptcy indicators among different business fields.