Agenda

Eshagh Jahangiri - Low-and-High-beta Stock Pair-trading by Reinforcement Learning Policies 

Abstract:  This study uses three Reinforcement Learning (RL) methods to implement  pair trading strategy considering one high-volatility and one low-volatility stock in the financial portfolio. The algorithms manage a two-asset financial portfolio in the New York stock exchange market (NYSE). The RL algorithms search for the optimal strategy to trade without having an accurate financial market model. A two-asset portfolio is used as a test-bed for applying RL methods based on function approximation, namely SARSA, Q-learning, and Greedy-GQ. To bring our trading system closer to the real one, we take into account a fixed transaction cost. In particular, we consider 648 combinations of different hyper-parameters in implementing of each RL algorithm. In most of the configurations our model performed better than the benchmark, the equally weighted portfolio, during the period from the 1st of October 2021 to the 1st of October 2022. Greedy-GQ algorithm achieved the best result among other two algorithms. We evaluated the impact of having more elements in the state vector. As a robustness test, we also ran our model during the 2007–2008 financial crisis, and it on average outperformed the benchmark.

The seminar can be attended also remotely, connecting to ZOOM: unive.zoom.us/j/83316158636
Meeting ID: 833 1615 8636