This study investigates the role of fairness in machine learning models for loan approval prediction, with a focus on racial equity. Currently, financial institutions process data for loan approval using a mix of technology and human oversight. In this approach, a Light Gradient Boosting Machine (LightGBM) was developed and compared against a baseline Logistic Regression model using loan data from the Home Mortgage Disclosure Act. LightGBM outperformed the baseline across predictive metrics, while also demonstrating stronger fairness outcomes. Fairness was assessed through demographic parity, equalized odds, and recourse burden gap, supported by counterfactual profile testing. Results show that the LightGBM model reduces racial disparities in loan approvals, with nearly balanced approval rates between privileged and unprivileged groups and lower recourse costs for disadvantaged applicants. These findings highlight the importance of integrating ethical considerations into credit risk modeling, showing that counterfactual analysis can improve both predictive performance and fairness.