Main shocks are calculated to cast stress shadows across broad areas where aftershocks occur. Thus, a key problem with stress-based operational forecasts is that they can badly underestimate aftershock occurrence in the shadows. We examine the performance of two physics-based earthquake forecast models (Coulomb rate/state (CRS)) based on Coulomb stress changes and a rate-and-state friction law for their predictive power on the 1989 Mw = 6.9 Loma Prieta aftershock sequence. The CRS-1 model considers the stress perturbations associated with the main shock rupture only, whereas CRS-2 uses an updated stress field with stresses imparted by M ≥ 3.5 aftershocks. Including secondary triggering effects slightly improves predictability, but physics-based models still underestimate aftershock rates in locations of initial negative stress changes. Furthermore, CRS-2 does not explain aftershock occurrence where secondary stress changes enhance the initial stress shadow. Predicting earthquake occurrence in calculated stress shadow zones remains a challenge for stress-based forecasts, and additional triggering mechanisms must be invoked.