In this paper we use a real-coding Genetic Algorithm to evolve three different architectures of Fuzzy Wavelet Neural Networks (FWNNs) and to compare their performances when predicting financial time series (stock exchange indexes) by means of NARX( ) models. FWNNs are designed by hybridization of Neural Networks, Fuzzy Inference Systems and Wavelet Functions as substitutes of Sigmoidal Activation Functions. The following FWNN architectures are evolved and compared: a summation-type FWNN, a multiplication-type FWNN and an extended version of the latter with both nonlinear and linear terms in the consequent part. Hybridization can be viewed as a way to overcome the limitations of individual techniques and to combine their benefits. It allows obtaining networks with the explanatory capabilities of fuzzy systems, a small number of neurons in the hidden layer, a fast training speed and a more precise and quick convergence. We use Matlab for implementing the algorithms and validate their predictive performance on data from Bucharest Stock Exchange.