Regression models for the prediction of grain yields of non-irrigated winter wheat in a semi-arid Mediterranean environment were developed based on drought variability. Twenty-five years (1980–2004) of climate data and yield data from four soils (sandy loam, clay, clay loam and sandy clay loam soil) in northern Greece were used for this purpose. Two variables were selected as explanatory variables of the models: (a) the monthly precipitation versus the monthly reference evapotranspiration ratio (P/ETo), which describes the monthly drought and consequently the water deficit conditions during the wheat-growing season and (b) the mean observed yield (y) of each soil, which indirectly describes the intrinsic fertility of the soils. A resampling technique using subsets of the data (bootstrapping) was applied to estimate the coefficients of the models, to assess the uncertainty of the selected explanatory variables and to validate the models. The models showed adequate predictive ability of wheat yields, defining the time and intensity of drought effects. The most crucial period for winter wheat was found to be primarily the vegetative-reproductive stage period between late winter and mid-spring (i.e. February to April). Soil clay content was found to be the most representative parameter in describing most of the physico-chemical parameters and properties of the soils and consequently the mean yield, indicating that yield is non-linearly correlated with most soil properties. With the proposed models, yield gap (YG) predictions between two growing seasons of the selected soils presented 84% accuracy in all years in the identification of the correct signal (+ or −) of yield increase or decrease, respectively, and adequate performance in the prediction of the mean YG.