Using elements DuPont analysis to estimate the return on equity based on a panel data and extreme gradient boosting approach

Abstract

This thesis investigates whether the elements of the DuPont analysis, including Net Profit Margin, Asset Turnover, and Financial Leverage, can serve as independent explanatory variables for estimating Return on Equity. To provide a broader understanding of the effectiveness of these elements ROE estimation was also conducted using lagged ROE and raw financial data for comparison with the DuPont-based model. A panel dataset is used based on firms listed in the S&P 500, the thesis applies both traditional econometric methods (Panel Data Analysis) and machine learning technique (XGBoost Regressor) to evaluate estimation performance. The findings reveal that Net Profit Margin, Asset Turnover, and Financial Leverage can each be effectively used to estimate ROE, with Net Profit Margin consistently identified as the most influential factor across both modeling techniques. Furthermore, the DuPont-based model outperforms the model using raw financial data or lagged ROE, achieving higher accuracy and interpretability, especially in the XGBoost framework. While the panel regression model offers theoretical transparency, the machine learning model demonstrates superior estimation power and better handles nonlinearity and data irregularities.