This paper introduces a novel data-driven option pricing methodology that relies exclusively on historical underlying asset prices, bypassing the need for option market data. The key idea is to recover a pricing kernel that transforms option pricing from the risk-neutral world to the objective world, where historical data is available. The pricing kernel is obtained by solving a logarithmic utility maximization problem, whose optimal wealth process is shown to be the reciprocal of the pricing kernel. This optimization is performed using deep learning, where portfolio strategies are parameterized by neural networks and trained on historical price paths. Once the pricing kernel is learned, option prices are computed by estimating conditional expectations under the objective measure, which is reformulated as a functional optimization problem minimizing a martingale loss function. The method ensures arbitrage-free prices and requires only a single training phase for all derivatives on the same underlying. Numerical experiments confirm the effectiveness of the approach, making it particularly useful for emerging markets with limited option data.