This study discusses the evolution of KSE-100 index returns as a dynamical system. We present application of nonlinear time-series analysis. Our results show that estimation of correlation dimension for the case of KSE-100 index returns is not possible. We further go into nonlinear analysis and construct a model of the series based on feedforward neural network with backpropagation training. We construct many neural networks and the one with Levenberg-Marquardt backpropagation is found to give slightly better results compared to ARMA/ARIMA models. Neural networks are found to be applicable in those cases when nonlinear time-series analysis is at failure.