This project applies a multi-method feature selection ensemble combined with a Multi-Layer Perceptron (MLP) to forecast cumulative stock returns across multiple horizons. A diverse set of technical indicators is processed through both descriptive-statistic-based and model-based feature selection methods, including variance filtering, Pearson correlation, LASSO, tree-based importance, dispersion ratio, and others. The selected features are used to train MLP models, whose performance is benchmarked against naive and persistent return forecasts. The feature selection ensemble significantly enhances MLP performance compared to using all features, though the naive benchmark remains difficult to outperform, reflecting the low signal-to-noise ratio typical in financial returns.