Adaptive Uncertainty Filtering for Automated Forex Trading Using LSTM and Conformal Prediction

Automated forex trading systems generally rely on machine learning models that generate point predictions without providing any measure of prediction confidence. Without this information, trading decisions made during volatile market periods carry higher risk of producing losses. This study proposes an Adaptive Uncertainty Filtering approach for automated forex trading by combining Long Short-Term Memory (LSTM) with Conformal Prediction (CP). The LSTM model generates price forecasts for the XAU/USD pair, while the CP layer produces prediction intervals that represent the degree of uncertainty in each forecast. Rather than applying a fixed uncertainty threshold as used in conventional approaches, the proposed method computes a dynamic threshold based on both the current prediction interval width and the recent average interval width. The system was implemented on the MetaTrader 5 platform and evaluated using live XAU/USD market data. The experimental results show that the adaptive filtering strategy reduces unnecessary trade executions during high-uncertainty periods and improves the overall win rate compared to the static threshold baseline. In future work, this framework can be extended to additional currency pairs and enhanced with alternative deep learning architectures.