Comparing methods to compute stock intrinsic values

This study systematically evaluates the accuracy and applicability of four mainstream valuation models—DCF, RIM, P/E, and P/B—using a sample of 3,070 Chinese A-share listed companies from 2015 to 2020. The empirical results demonstrate that relative valuation methods, particularly the P/B model, exhibit superior accuracy and stability compared to absolute valuation models within the volatile A-share environment. Quantitatively, the P/B model yields the lowest mean valuation bias (0.691) and highest explanatory power (R²=0.692), followed by the P/E model (mean bias=1.003, R²=0.435) and RIM model (mean bias=1.285, R²=0.489), while the DCF model shows the poorest performance (mean bias=1.677, R²=0.037). By constructing and employing a diagnostic machine-learning-based valuation framework—integrating ensemble algorithms such as Random Forest, XGBoost, LightGBM, and CatBoost—the research reveals that the success of the P/B model stems from its effective anchoring of high-weight fundamental signals (net assets). Conversely, the failure of the DCF model is primarily attributed to the inability of its linear framework to capture complex non-linear interactions between financial indicators. Furthermore, the study identifies significant industry heterogeneity, suggesting that valuation strategies must be dynamically adjusted based on specific industrial characteristics. This research provides a novel "theory-driven + algorithmic diagnosis" paradigm for understanding valuation disparities in complex capital markets.