Bridging Prediction and Profit: Deep Learning models with Trading Evaluation for FTSE 100

Abstract

This study examines stock price forecasting for FTSE 100 companies using deep learning and XAIl. The research addresses the disconnect between predictive accuracy and interpret ability in financial models by integrating data-driven forecasting with transparent feature attribution. Four neural architectures: LSTM2, Bi-LSTM3, GRU4 and CNN5 are compared to classical benchmarks: SMA6 and EMA7. Models are trained on OHLCV8 data augmented with technical indicators. Evaluation uses a threshold-based trading strategy. The findings indicate that a lower prediction error does not necessarily result in higher profitability. Although LSTM achieved the lowest prediction error, GRU and Bi-LSTM produced more stable cumulative returns (16%), compared to the EMA benchmark (2%). SHAP9 analysis demonstrates that recent price movements and momentum indicators, particularly SMA, drive model decisions.