The hypothesis in this paper is that a significant amount of intraday market data is either noise or redundant, and that if it is eliminated, then predictive models built using the remaining intraday data will be more accurate. To test this hypothesis, we use an evolutionary method (called Evolutionary Data Selection, EDS) to selectively remove out portions of training data that is to be made available to an intraday market predictor. After performing experiments in which data-selected and non-data-selected versions of the same predictive models are compared, it is shown that EDS is effective and does indeed boost predictor accuracy. It is also shown in the paper that building multiple models using EDS and placing them into an ensemble further increases performance. The datasets for evaluation are large intraday forex time series, specifically series from the EUR/USD, the USD/JPY and the EUR/JPY markets, and predictive models for two primary tasks per market are built: intraday return prediction and intraday volatility prediction.