Lighting Up Dark Pools

Abstract

In recent years, dark pools have become an increasingly popular trading venue, especially for high-frequency traders. However, this increase in popularity has not been without controversy. In this thesis, we explore the value these unique venues bring to traders through the study of the dark pool problem. Here an agent faces the decision of placing orders to a set of dark pools over time in order to achieve a specified total execution size in each time step. While doing this, the agent tries to learn the distribution of the amount of liquidity available at each distinct dark pool in order to determine the optimal allocation of orders. We extend previous work by Ganchev et al. and Agarwal et al. from 2010 by adding lit markets as a trading option for the agent, bringing the model closer to a realistic setting. To solve this extended problem, we build on previous results and develop an algorithm for the new setting. We evaluate the existing and new algorithms through both theoretical and simulated means and find that the algorithm of Agarwal et al. outperforms that of Ganchev et al. in our simulated environment. Next, we modify the dark pool problem by making the assumption that each liquidity distribution is Poisson, and attempt to learn the parameter for each pool. To do this, we introduce a maximum likelihood estimator for censored Poisson data, and simulate its use in the dark pool problem. Through our simulations, we find instances where dark pools provide valuable opportunities for traders to avoid market impact. This lends credence to regulatory efforts that ensure fairness in dark pools, rather than those that restrict their use.