Business Valuation with Machine Learning

Abstract

The aim of the article is to test the hypothesis about the applicability of machine learning methods to train models that allow to accurately predict the market capitalization of an enterprise based on data contained in three main forms of financial statements: Income statement, Balance sheet, and Cash flow statement. The scientific novelty of the study lies in the proposal of an alternative approach to the actual finance problem — business valuation. The conducted empirical study allows us to test the hypothesis under consideration. We train various models using the most popular machine learning methods (LASSO, Elastic Net, KNN, Random Forest, SVM, and others). To determine the best approach for assessing the value of a company, the effectiveness of different methods is compared based on the R2 performance metric (86,7% for the GBDT). Financial statements data of NYSE and NASDAQ companies are used. The study also addresses the problem of the interpretability of the trained models. The most important features are identified — the forms of financial statements and their specific items that have the greatest impact on market capitalization. Three independent ways to determine feature importance indicate the significance of the information contained in the Income statement. In particular, Comprehensive income was the most important item for accurate predictions. Robust methods of variable normalization and missing data imputation are also highlighted. Finally, various ways of improving the developed models are recommended to achieve even higher accuracy of forecasts. The study concludes that machine learning can be applied as a more accurate, unbiased, and less costly approach to value a company. Feature importance analysis can also be used to understand and further explore the value creation process.