Automated identification of fraudulent financial statements by analyzing data traces

Firms are obliged by law to publish accurate financial statements. Nevertheless, cases exist where publicly issued documents hide the actual bad financial statuses of companies and this is revealed years later. Since companies publish financial figures periodically, it is interesting to examine whether monitoring those values or ratios based on them can help in early detection of fraud. In this work, a dataset was constructed including 943 firm-years of Greek companies enlisted at the Athens Stock Exchange, for the period 2005-2018. Experiments using combinations of financial ratios and various machine learning algorithms were undertaken in an effort to construct a system able to automatically detect false and misleading financial statements, that were issued legitimately by firms. Several instantiations of a machine learning workflow were tested using various classifier algorithms including AdaBoost, Random Forests, and others. Experiments showed that companies that issue false financial statements can be spotted automatically for most of the cases, years before the problem is manifested. So, the potential of early detection of seemingly healthy, but in fact, distressed companies do exist. An automated tool can be constructed that should be useful for financial analysts, investors and the capital markets authorities.