We document bi-modality of relative stock returns when conditioned on momentum factors, which makes the momentum strategy fundamentally risky. Using a deep neural network with momentum factors as input, we estimate cross-sectional return distributions and use them to sort stocks based on their predicted financial performance. Our method alleviates the bi-modality of stock returns and renders significantly improved portfolio performance. Tested on the US market, an equal-weighted long-short strategy yields a Sharpe ratio as high as 2.6 over the period from 1975.01 to 2017.01. When we add a size variable to enhance the performance of a value-weighted portfolio, we achieve a Sharpe ratio of 2.9 for an equal-weighted portfolio and 1.9 for a value-weighted portfolio. Remarkably, momentum crash disappears in our model without any special consideration.