Using high-quality hydraulic conductivity test results for compacted soil samples with widely different gradation characteristics, 14 commonly employed empirical correlations for estimating the permeability coefficient (k) were evaluated. Comparisons of measured and predicted k values indicated most predictions were within a margin of ±100%, although they could vary by up to 500% or more. Statistical analysis indicated that high coefficients of determination obtained for these datasets can be grossly misleading, since the data may have markedly high values of the standard error of the estimate. Grain size characteristics and void ratio were identified as the most sensitive input parameters for the models investigated. On this basis and using the results of hydraulic conductivity testing performed on 20 soil samples with very different gradations, a new prediction model that expressly considers both gradation and void ratio is proposed. Test results from an additional 16 sandy soil samples were used to validate the new model, which was found to produce a fairly good distribution of data points within 95% prediction intervals and an equitable spread about the unity line in a plot of predicted against measured k values. The experimental data were also analysed using the grading entropy framework.