News Sentiment Analysis in Forex Trading Using R-CNN on Deep Recurrent Q-Network

Abstract

Every trader in trading aspires to make the best decisions in buying and selling transactions and maximize the profits they get. The reinforcement learning method is a growing and popular method for making predictions in financial markets. After the AlphaGo defeated the strongest Go Contemporary board game player named Lee Sedol in 2016, this method creates a system capable of learning trading from itself. In a systematic review conducted by Terry Lingze Meng, all the latest articles related to stock and forex predictions that use reinforcement learning as the primary method only use past technical data as their state. In this study, the authors propose the implementation of word2vec and Recurrent Convolution Neural Network to provide the agent with the ability to read and process fundamental factors through the provided news headlines. The action augmentation technique reduces random exploration by the agent. The simulation will run on historical price changes for the seven most frequently traded currency pairs. This implementation demonstrates the impact of adding news headlines to improve risk management and lower the maximum withdrawal point value on almost all tested currency pairs with the highest increase of up to 57.9% on GBPUSD from 7.9% to 3.32%.