Firm valuation using accounting-based capital structure and cash holdings: An explainable machine learning approach

This study investigates the impact of cash holdings and capital structure on firm valuation in Egypt's emerging market, examining how COVID-19 altered investor perceptions. The research employs explainable machine learning (ML) to uncover non-linear financial thresholds that traditional valuation models overlook. Egyptian listed firms from 2015 to 2022 are analyzed using a Super Learner ensemble (Extremely Randomized Trees, Extreme Gradient Boosting, and a Linear Regression meta-learner) alongside SHapley Additive exPlanations (SHAP) and partial dependence analysis, with the Super Learner's performance compared against conventional methods in assessing financial policy effects on Tobin's Q. Three key findings emerge: (1) Leverage exhibits a non-linear relationship with valuation, where extreme levels (LEV > 1.2) unexpectedly enhance firm value, challenging trade-off theory; (2) Cash holdings demonstrate threshold effects, with optimal value at ~40% of assets and sharply increasing marginal benefits beyond this point; and (3) COVID-19 amplified these dynamics, elevating the liquidity premium while penalizing excessive debt. The Super Learner significantly outperformed traditional statistical and ML models (R² = 0.572 vs. 0.19-0.47). Practical implications suggest that investors and managers in emerging markets should adopt dynamic cash-debt optimization to avoid undervaluation, while policymakers can use ML-driven thresholds to design crisis-responsive regulations. This study contributes to the literature by (1) identifying non-linear thresholds that extend trade-off and pecking order theories, (2) introducing explainable ML to valuation research to balance accuracy and interpretability, and (3) providing novel evidence of COVID-19's structural impact on investor behavior in emerging economies.