Modelling of shear behaviour of residual soils is difficult in that there is a significant variability in constituents and structures of the soil. A Recurrent Neural Network (RNN) is developed for modelling shear behaviour of the residual soil. The RNN model appears very effective in modelling complex soil shear behaviour, due to its feedback connections from an hidden layer to an input layer. Two architectures of the RNN model are designed for training different sets of experimental data which include strain-controlled undrained tests and stress-controlled drained tests performed on a residual Hawaiian volcanic soil. A dynamic gradient descent learning algorithm is used to train the network. By training only part of the experimental data the network establishes neural connections between stress and strain relations. Although the soil exhibited significant variations in terms of shearing behaviour, the RNN model displays a strong capability in capturing these variabilities. Both softening and hardening characteristics of the soil are well represented by the RNN model. Isotropic and anisotropic consolidation conditions are precisely reflected by the RNN model. In undrained tests, pore water pressure responses at various loading stages are simultaneously simulated. With a RNN model designed for a special drained test, the network is able to capture abrupt changes in axial and volumetric strains during shearing courses. These good agreements between the measured data and the modelling results demonstrate the desired capability of the RNN model in representing a soil behaviour. 1998 John Wiley & Sons, Ltd.