This project determined whether shallow or deep artificial neural networks performed better for noisy signal forecasting. Using a shared convolutional neural network feature extractor, we compared the performance of similar fully connected networks with varying depth, from 1 layer to 10 layers, when forecasting signals with predetermined signal to noise ratios: {0.25, 0.5, 1, 2, 3, 4}. We averaged loss results (MSE loss) for each model depth and found that shallower networks outperformed deeper ones, with optimal depth being under 5 layers.