Journal of Robotics, Automation and Smart Systems

ABSTRACT

Predicting stock prices remains one of the most challenging problems in applied machine learning not because markets are random, but because the structure underlying price movements is continuously contested and revised by millions of participants. Rather than pursuing a marginally better single model, this paper presents an Agentic AI system that assembles a coordinated team of specialized agents, each responsible for a distinct information channel, collaborating to produce trading decisions. Four agents a Market Agent reading price momentum and volatility, a Sentiment Agent running FinBERT on financial news, a Prediction Agent backed by a trained LSTM, and a Decision Agent resolving disagreements through majority vote-feed into a dynamic inverse-RMSE ensemble of LSTM, XGBoost, and Transformer models. A Q-learning reinforcement learning layer then refines the final decision based on realized price consequences. The system is evaluated on nine years of Apple Inc. (AAPL) closing price data spanning 2015 to 2024. The LSTM achieves RMSE = 5.44 USD, MAE = 4.64 USD, and R² = 0.915. The dynamic ensemble (weights: LSTM 0.625, XGBoost 0.338, Transformer 0.037) improves R² to approximately 0.96. Backtesting a $10,000 portfolio yields a CAGR of 8.69%, a Sharpe Ratio of 0.65, a maximum drawdown of −10.04%, and a final portfolio value of $11,566.98.

Keywords: Agentic AI; Multi-Agent Systems; LSTM; Transformer; XGBoost; FinBERT; Reinforcement Learning; Algorithmic Trading; Stock Market Prediction; Ensemble Learning