Multi Q-Learning: Ensemble Reinforcement Learning for Stock Market Trading

Description:
The stock market, based on a large number of stocks of a huge range of different com- panies, is highly dynamic and complex. Its dynamic is based on hard data i.e. stock price history, sales or earnings and soft data i.e. opinions, psychiological factors, mar- keting or sentiment. Due to this complexity it is nearly impossible to make reasonable predictions of the future market behaviour without powerful computer algorithms. Aim of this work is to develop different Q-learning trading agents and build a trading ensem- ble model from these agents. For building the different Q-learning agents used within the ensemble, different machine learning regression methods are implemented to replace the Q-table. These trading agents perform trading decisions on historic price data of S&P-500 and DAX. On top of that, in order to test the performance of the built trading agents and the ensemble model on as different performing stocks as possible, unsuper- vised learning clustering methods are used to seperate different stock clusters and the agents trading performances are tested on stocks of each cluster.