Water repellency (WR) is a common soil property in many fire-affected ecosystems, but it also occurs in long-unburned terrain. It can vary in space at different scales (between point and pedon or slope and catchment) and time (during the same day, between seasons or years, or with a post-fire recovery period). This paper: i) reports on the occurrence and persistence of WR in fire-affected calcareous forest soils under Mediterranean climatic conditions, examining its spatial variability at macro-, meso- and micro-scales, and monthly changes with soil moisture content; and ii) develops exploratory models to estimate the probability of the natural background (not fire-induced) WR to occur through a Mixed-Effect Logistic Regression Model. Four sites with comparable soil and vegetation types were studied, all of them burned in 1979, the third again 1999, and the fourth in April 2008. All sites were sampled immediately after the fire of 2008, a further 17 times until July 2009, and once more in August 2011. At each site, 5 random plots (10 cm × 10 cm) were selected for each of the vegetation types: Pinus halepensis, Quercus coccifera, Rosmarinus officinalis and bare soil. In each plot we carried out 10 Water Drop Penetration Time measurements at the surface and at 1 cm depth. WR was detected in samples from all sites, although WR was most frequent at the soil surface at the sites last burnt in 1979. The recently burned site had similar WR to the long-unburned ones in surface but greater at 1 cm depth, although for both depths WR was reduced by the following year. WR was still very low in the site burned 10 years before. Variability of WR between different vegetation types was as high as within the same type at the same site, and similarly high at within sites (1 ha) and at 10 cm × 10 cm plot-scales. The lowest variability was found in bare soil plots because they were mostly wettable. An exploratory model to estimate the probability of WR occurrence was derived for each unburned site. The most powerful explanatory variables for the probability of WR to occur (P < 0.0001) for both unburned sites were the covariate moisture content together with the category soil surface (associated to the fixed factor soil depth) for one site and P. halepensis (associated to vegetation type) for the other. The model input parameters are straightforward to obtain and the model may be a useful tool in estimating occurrence and fluctuation of soil WR.