Enhancing forex market forecasting with feature-augmented multivariate LSTM models using real-time data

This study proposes a feature-augmented multivariate LSTM model for real-time Forex market forecasting. By incorporating engineered financial indicators—such as Close_Change, RSI, and gold price—alongside traditional OHLCV data, the model captures nonlinear temporal dynamics and macro-financial interactions. A sliding window approach structures input sequences for a stacked LSTM network optimized for short-term prediction. Experimental results on major currency pairs demonstrate that the proposed model outperforms baseline LSTM, GRU, and classical machine learning methods in RMSE, MAE, and MAPE metrics. Statistical validation using the Wilcoxon signed-rank test confirms the improvements are significant. The model's robustness under volatility stress and noisy inputs highlights its practical relevance for real-time decision-making. Potential extensions include incorporating news-based sentiment and multimodal signals to enhance adaptability.