Maximizing REITurns: A Multi-Armed Bandit Approach to Optimizing Trading Strategies on Real Estate Investment Trusts

Abstract

The overarching goal of this thesis was to research and develop an effective trading strategy specific to REIT stocks. Real Estate Investment Trusts, commonly referred to as REITs, are companies that own incoming generating real estate properties. REIT stocks are common investments for investors that are seeking larger dividends and therefore do not expect high growth. The topic of REIT stock movement and the application of trading strategies in REIT stocks has been relatively under explored in prior academic research. These topics have been traditionally studied in common stocks, however, the REIT market provides interesting qualities that may have been overlooked in the past. In this paper, we will experiment with applying reinforcement learning techniques to optimize the use of traditional trading algorithms on REIT stocks. To do this, we research and apply methods from Multi-Armed Bandit problems. We set a Multi-Armed Bandit problem where the arms exist as trading strategies and the rewards of each arm are defined as the returns from that trading strategy. The most complex issue in transforming this into a Multi-Armed Bandit problem was addressing the non-stationarity of returns in stocks. After a thorough analysis, we focus on two algorithms that address the seasonality of returns in each trading strategy. We develop multiple automated trading models and we apply these models to sets of historical data to determine if the Multi-Armed Bandit algorithms carry any efficacy in REIT stock trading.