There are a number of models that are used for mass appraisal of properties. However, the choice of a model is predicated on a number of criteria. One of these criteria is to compare models predictive accuracies that are reflected in minimum error of estimates. This study focuses on comparing predictive accuracies of mass appraisal models with a datasets of 3494 property transactions from the city of Cape Town, South Africa. Five mass appraisal models including back propagation trained artificial neural networks, multiple regression model, M5P tree, support vector machine optimise with sequential minimal optimisation and additive nonparametric regression were used for the simulations. Waikato Environment for Knowledge Analysis (WEKA) explorer; an open source data mining software was used to pre-processed property data to normalised values and model property prices. The analysis shows that BP trained artificial neural networks (BP-ANN) and M5P tree utilised in this study predicted better results with root mean squared error and mean absolute error within acceptable threshold of 5%. But M5P tree shows distinctiveness in predicted results between normalised and absolute values which require further examination. The other three mass appraisal models including multiple regression model, additive nonparametric regression and support vector machines with simulated minimal optimisation predicted RMSE that are higher than 5% acceptable threshold. With these results it is hereby recommended that mortgage lenders, valuation offices in South Africa, the rest of Africa and beyond should consider utilising BP-ANN in their mass appraisal predictions.