A Multi-LSTM Ensemble and PPO-Driven Trading Agent with Integrated FinBERT Sentiment Intelligence

Abstract

Financial market is a non-linear dynamical system which is very complicated, which is conditioned by a great number of factors, starting with historical price, tendencies towards abrupt changes of world opinion [1]. Traditional algorithmic trading tends to fail since it is based on fixed rules that are not adaptive to the stochastic character of the mar-ket. [4] [7] To solve this, we would suggest FinSynapse and FinBERT Sentiment Intelligence, an improved trading agent that combines Deep Reinforcement Learning. Natural Language Processing (NLP) (DRA) [7] [12]. Our hybrid architecture which works on by implementing group of parallel LSTM Networks to dis-extract time characteristics like momentum, trend and volatility which is generated through a Proximal Policy Optimization (PPO) agent [11] and it is used for decision making. Moreover, we combine FinBERT to quantify financial news market sentiment, which can be adjusted by the agent for decision making. Backtesting on Dow Jones data indicates that this mixed strategy works with a Accumulative Return of 97.44 and Sharpe Ratio of 2.21, significantly becoming superior to conventional baseline models.