A Data-Driven Deep Learning Approach for Options Market Making

We develop a data-driven approach for options market making. Using stock options data from CBOE, we find that both buy and sell orders exhibit strong self-excitation but insignificant cross-excitation. We show that a Hawkes process with a time-varying baseline intensity and the power law kernel provides a good fit to the data of market order flows for stock options. To solve the optimal market making problem for a single option, we approximate the market making strategy at each decision time by a neural network and train them to optimize the expected utility of the market maker. We study feature selection for the neural networks and compare the out-of-sample performance of the optimal neural network strategy trained from data generated by the Hawkes process and the Poisson process. We find that using the more realistic Hawkes model improves the out-of-sample performance significantly. Furthermore, utilizing the Hawkes process intensity or the expected number of market order arrivals computed under the Hawkes model as an additional input feature can boost the performance. We also show how to solve the market making problem for option portfolios with Greeks and inventory constraints using neural network approximation