Pedotransfer functions (PTFs) are becoming a more common way to predict soil hydraulic properties from soil texture, bulk density, and organic matter content. Thus far, the calibration and validation of PTFs has been hampered by a lack of suitable databases. In this paper we employed three databases (RAWLS, AHUJA, and UNSODA) to evaluate the accuracy and uncertainty of neural network-based PTFs. Sand, silt, and clay percentages and bulk density were used as input for the PTFs, which subsequently provided retention parameters and saturated hydraulic conductivity, Ks as output. Calibration and validation of PTFs were carried out on independent samples from the same database through combination with the bootstrap method. This method also yielded the possibility of calculating uncertainty estimates of predicted hydraulic parameters. Calibration and validation results showed that water retention could be predicted with a root mean square residual (RMSR) between 0.06 and 0.10 cm3 cm−3; the RMSR of log(Ks) was between 0.4 and 0.7 log (cm day−1). Cross-validation was used to test how well PTFs that were calibrated for one database could predict the hydraulic properties of the other two databases. The results showed that systematically different predictions were made when the RMSR values increased to between 0.08 and 0.13 cm3 cm−3 for water retention and to between 0.6 and 0.9 log(cm day−1) for log(Ks). The uncertainty in predicted Ks was one-half to one order of magnitude, whereas predicted water retention points had an uncertainty of about 0.04 to 0.10 cm3 cm−3. Uncertainties became somewhat smaller if the PTFs were calibrated on all available data. We conclude that the performance of PTFs may depend strongly on the data that were used for calibration and evaluation.