The paper presents the results of investigating the strength characteristics of eleven compositions of protective and decorative epoxy polymer coatings with mineral fillers and an unfilled polymer under natural climatic aging in a temperate climate (Saransk). The work formalises the methodological approach to the formation of a training dataset for predicting changes in the strength characteristics of polymer composites with mineral fillers using machine learning methods. It is shown that with the introduction of a mineral filler, the failure mode of the epoxy matrix under bending changes from viscous-fluid to brittle, which allows the bending strength to be unambiguously determined. Under uniaxial compression, the classical strength value is reached in the region of relative strains of 50 % and higher, which corresponds to the densification phase of the already failed material and does not physically reflect the actual strength response. As a methodologically correct alternative, the yield strength is proposed, identified as the first local maximum on the smoothed σ–ε curve and unambiguously determined for all studied compositions at all natural climatic aging points. The relative difference Δσ between the strength value (in bending) or the yield strength (in compression) of the filled composition and the corresponding stress value on the deformation curve of the unfilled polymer, taken at the same point of relative strains for each time point of climatic aging, is proposed as the target feature of the prediction model. This approach automatically eliminates from the target variable the components associated with the post-curing effects and the degradation of the polymer matrix itself, leaving the model with the task of predicting the contribution of the filler and the state of the polymer–filler interface. Based on the analysis of the reinforcing effect of the eleven filled compositions in the reference state without post-curing, three groups of fillers are identified according to the level of the relative strength gain under bending («low» 3÷10 %, «moderate» 24÷53 %, «high» 70÷120 %), which are consistent with the physical picture of the filler–polymer matrix interaction. Recommendations are formulated on the composition of the model's input feature space, including the type and shape of the filler particles, the mass fraction, the characteristic particle size, the specific surface area, and the characteristic relative strain of the composition in the reference state. It is noted that the bending tests possess substantially greater informativeness in the assessment of the climatic resistance of the investigated class of materials in comparison with uniaxial compression.