Machine learning in firm performance analysis: Regularization methods

In this study, it is aimed to compare linear regression models estimated by regularization methods (ridge, lasso and elastic-net) and ordinary variable selection methods that use ordinary least squares estimator in terms of the variable selection and prediction performances on a real business problem. For the comparative analysis the effect of financial ratios on firm performance is analyzed. Thus, this thesis also seeks to determine which liquidity, debt and efficiency ratios affect the firm performance and to present the mathematical structure of this relationship. Manufacturing firms listed and traded in Borsa İstanbul during the 2012-2019 period are sampled for the analysis and among profitability ratios return on assets (ROA) and return on equity (ROE) are chosen as proxies of firm performance. According to the findings, lasso and elastic-net methods have more stable behavior on variable selection. The model with the highest estimation success is the model whose variables are selected by lasso and then the coefficients are estimated by ordinary least squares. Also, the asset turnover ratio and leverage ratios affect the return on asset the most. On the other hand, total assets/equity and bank credits/total assets ratios are the most influential financial ratios over return on equity. In conclusion, manufacturing firms should decrease their dependency on their debt and keep their stock and long-term assets at minimum in order to enhance their firm performance.