Abstract
We introduce a novel approach to options trading strategies using a highlyscalable and data-driven machine learning algorithm. In contrast to traditionalapproaches that often require specifications of underlying market dynamics orassumptions on an option pricing model, our models depart fundamentally fromthe need for these prerequisites, directly learning non-trivial mappings frommarket data to optimal trading signals. Backtesting on more than a decade ofoption contracts for equities listed on the S&P 100, we demonstrate that deeplearning models trained according to our end-to-end approach exhibitsignificant improvements in risk-adjusted performance over existing rules-basedtrading strategies. We find that incorporating turnover regularization into themodels leads to further performance enhancements at prohibitively high levelsof transaction costs.