Master Thesis

Abstract

The digitalisation of financial markets enabled the rise of algorithmic trading in market making. The role of a market maker is to maximise their profit and loss (PnL) by posting optimal bid and ask quotes to capture the spread whilst minimising the inventory and adverse selection risk. Analytical optimal market making approximations are predicated on a set of assumptions about the underlying market dynamics. Therefore, their results, in the best case, only apply partially to the real world of financial markets. Model-free deep reinforcement learning (DRL) is successfully being applied to difficult sequential decision-making problems in many areas such as games and robotics. Recent market making publications find model-free DRL advantageous since it neither requires explicit modelling of the underlying market dynamics nor any prior knowledge of optimal market making. Thus, market making DRL agents tend to be designed to ingest numerous features which can originate from other machine learning models. In this thesis, we apply the latest DRL algorithm to market making, focusing on well established Order Book Imbalance (OBI) and the recently introduced Book Exhaustion Rate (BER) to mitigate directional risk. Experimental results show that even the simplest inventory DRL agent successfully learns a near optimal market making strategy, by training on as little as one day of data. Moreover, the two limit order book signals OBI and BER help to further improve the performance and mitigate the directional risk a market maker faces. More specifically, we achieved 10% higher terminal wealth and just 32% of the maximum drawdown of the benchmarks.