Research by Walczak and Cerpa [32] and others [27, 39] has already focused on the question of selecting appropriate input values for modeling time series domains. However, the question of how much historical information is required to produce the best-performing model has not been addressed by neural network researchers. The research presented in this article investigates the requirements for training or modeling knowledge when building nonlinear financial time series forecasting models with neural networks. Homogeneous neural network forecasting models are developed for trading the U.S. dollar against various other foreign currencies. The differences between the neural network models for a specific currency lie solely in the quantity of training data used to develop each time series forecasting model