Knowledge of soil hydraulic and solute transport properties, especially the susceptibility to preferential flow and transport is becoming increasingly important for managing our environmental resources. Such knowledge cannot be acquired by direct measurements at the field scale or larger for practical reasons. Instead, soil hydraulic and solute transport properties must be inferred from proxy variables that are easier to obtain. Potentially useful proxy variables include soil properties like texture and organic carbon content, site factors like land use or land management, hydrologic initial and boundary conditions as well as the scale of the investigated domain. Such an approach is commonly referred to as pedotransfer function. During recent years, we have assembled a large meta-database of solute transport experiments on undisturbed soil columns with inert tracers from the refereed literature (N=560). In this study, we used random forests to build pedotransfer functions from the datasets in the meta-database for predicting the relative 5%-arrival time which is an indicator for preferential transport. Using a random forest has the advantage that it not only allows for a prediction of soil susceptibility to preferential transport but the importance of individual proxy variables used in the prediction can also be quantified. Our results show that a random forest predicts the relative 5%-arrival time to a high accuracy with 10-fold cross-validation if breakthrough experiments from identical source-references were included in both the training and the validation set. If all datasets from specific source-references were excluded from the training set and only used for validation, the accuracy decreased noticeably. This suggests that the meta-database is, despite its considerable size, still too small to account for the diversity in soil properties, site factors, scale of experiment, and experimental setups. Furthermore, we found that soil texture was the most important predictor for preferential transport, which only occurred for sand contents below 90% and clay and silt contents above 10%. The next-most important proxy variables were the hydrologic initial and boundary conditions and the scale of the experiment.