Detection of Corporate Fraud Using Textual Analysis and Machine Learning: Evidence From Russian Companies

This paper is devoted to the detection of corporate fraud using textual data analysis and machine learning techniques on the example of Russian companies. The study involved the manual collection of English texts of annual reports, financial indicators from the Cbonds database, and information on corporate fraud from the press from 2014 to 2021 for 260 non-financial companies included in the ExpertRa ranking of Russia's largest companies (2021). The final sample, after removing missing data, includes a total of 386 observations for 55 companies over 8 years. The study used both variables of tone of the text and readability, commonly used in such papers, and variables of the vectorized text, which are rarely used, to analyze the messages of the company's management to shareholders. In addition, similar studies have not been conducted on a sample of Russian companies, which taken together make up the scientific significance of this study. The text tone variables used include the proportion of positive, negative, constraining, and uncertain words, calculated using the Loughran & McDonald dictionary. Text length and Fog-index were used as text readability variables. Vectorized words were obtained using the TF-IDF method. As a result, it was found that comparing to fraudulent companies, non-fraud companies, the texts of management messages to shareholders contained more positive words, fewer negative words, and were generally longer. In addition, it was proved that the detection of non-linear connections using machine learning methods can increase the predictive power of the models: among the models used (Logit, KNN, SVM, Decision Tree and Random Forest) was identified the optimal one - Random forest, the percentage of correctly predicted fraud cases for which was 75%, which is 2.5 higher than for the traditional logistic regression model. The optimal model was used to obtain word clouds indicating the presence or absence of fraud in the company. The model and word clouds obtained in this study can be used by forensic and audit specialists, government regulators, as well as by the company's stakeholders to make investment and other decisions.