Optimized Forex Trading using Ensemble of Deep Q-Learning Agents

Foreign exchange market is a fast paced market influenced by a wide variety of factors, which requires the careful study of data to take an advantageous position. There are many AI based approaches, chief among them being the use of single RL agent based systems trained on a specific time period. We have proposed the use of a Deep Q Learning based algorithm consisting of multiple agents trained(run) on different epochs on the dataset. Different feature extraction techniques like CNN, LSTM and their combinations have been tried to ascertain the better technique for our usecase. Decisions taken on each day are logged during validation and testing. An ensembling technique is used on these decisions to get the final decision and profit. It was observed that the combination of CNN and LSTM layers performed better than when each of the layers where tested in isolation from one another, howoever they result in overfitting. LSTM was found to be better overall and using an agreement threshold of 60% provided best results.