Using big data, and machine learning we show that investors can quite successfully predict delta-hedged (or delta-neutral straddle) option returns. The predictability is driven by option specific predictors, and commonly used stock predictors play only marginal role. Large set of stock predictors and machine learning also fail to robustly predict stock returns in more recent data. However, options-based predictors and machine learning allows identifying mispricing even for the cross-section of the most liquid S&P500 stocks. Moreover, machines almost unanimously identify options illiquidity as the main predictor of stock returns. Guided by machines, we uncover positive option illiquidity premium in the stock returns. Stock long-short value-weighted portfolio strategies formed on option illiquidity outperform machine learning based portfolios with Sharpe ratio of 2.