Enhanced Gold Price Prediction and Automated Trading Using FinBERT, CEEMDAN, Attention-LSTM and Proximal Policy Optimization

Abstract

Accurate gold price prediction and automated trading remain challenging due to the non-stationary, noisy nature of financial time series and the influence of macroeconomic factors. While Awad et al. demonstrated the effectiveness of combining NLP with a vanilla Deep Q-Network (DQN) trained on VMD-decomposed signals and BERT-based sentiment features, the long term viability of deep reinforcement learning (DRL) was not established. This paper proposes an improved framework by replacing these components with more robust alternatives: FinBERT for domain-specific sentiment analysis, CEEMDAN for adaptive signal decomposition, an Attention-LSTM trained with Huber loss for multi-step forecasting, and Proximal Policy Optimization (PPO) for more stable trading. The model also incorporates macroeconomic indicators (US Dollar Index and crude oil) and technical indicators (Bollinger Bands and MACD). Experiments on gold futures data (2012 to 2019) achieved an Annualized Sharpe Ratio of 62.4, Annualized Return of 1.34, and Calmar Ratio of 3.8, outperforming the DQN baseline. Ablation studies further confirmed the individual contributions of CEEMDAN and FinBERT.