We use machine learning to examine the predictability of hold-to-maturity returns on single stock options. In sharp contrast to the implications of standard asset pricing theory, we find that the expected return of the underlying stock fails to forecast option returns, but does explain cross-sectional variation in option prices. Option trading strategies based on the underlying’s expected stock return deliver anomalously low returns. These findings challenge canonical option pricing models and suggest that options are not a suitable instrument to harvest stock risk premia.