This paper applies deep learning techniques to uncover novel return predictability in the cross-section of delta-hedged equity options. I demonstrate sizable profits in long-short option portfolios using a Convolutional Neural Network (CNN) which automatically extracts relevant patterns from the implied volatility surface. Portfolio returns remain statistically and economically significant even after accounting for transaction costs. The CNN subsumes some commonly used predictors and cannot be fully explained by option-based characteristics. I further show that the CNN generates abnormal returns compared to a latent factor model that is based on a broad range of option and stock characteristics. Finally, I provide evidence that the model can also be used to predict returns of alternative option positions such as straddles.