The inclusion of soil wetness information in empirical landslide prediction models was shown to improve the forecast goodness of regional landslide early warning systems (LEWSs). However, it is still unclear which source of information - numerical models or in situ measurements - is of higher value for this purpose. In this study, soil moisture dynamics at 133 grassland sites in Switzerland were simulated for the period of 1981 to 2019, using a physically based 1D soil moisture transfer model. A common parameterization set was defined for all sites, except for site-specific soil hydrological properties, and the model performance was assessed at a subset of 14 sites where in situ soil moisture measurements were available on the same plot. A previously developed statistical framework was applied to fit an empirical landslide forecast model, and receiver operating characteristic analysis (ROC) was used to assess the forecast goodness. To assess the sensitivity of the landslide forecasts, the statistical framework was applied to different model parameterizations, to various distances between simulation sites and landslides and to measured soil moisture from a subset of 35 sites for comparison with a measurement-based forecast model. We found that (i) simulated soil moisture is a skilful predictor for regional landslide activity, (ii) that it is sensitive to the formulation of the upper and lower boundary conditions, and (iii) that the information content is strongly distance dependent. Compared to a measurement-based landslide forecast model, the model-based forecast performs better as the homogenization of hydrological processes, and the site representation can lead to a better representation of triggering event conditions. However, it is limited in reproducing critical antecedent saturation conditions due to an inadequate representation of the long-term water storage.