Algorithmic Trading Boosted by Artificial Intelligence. An LSTM-based approach to Intraday Gold Trading Using Technical Indicators

Sammanfattning

Financial markets generate large volumes of time series data, making them a natural domain for the application of machine learning methods. In particular, deep learning models such as Long Short-Term Memory (LSTM) networks have shown promise in modelling temporal dependencies and forecasting time series. At the same time, technical indicators, such as Simple Moving Average (SMA), Relative Strength Index (RSI), Moving Average Convergence Divergence (MACD) and Bollinger Bands, remain widely used by practitioners to identify trading signals based on historical price patterns. The extent to which the incorporation of such technical indicators into deep learning frameworks can improve the performance of algorithmic trading is an active area of research. This thesis investigates the effectiveness of combining an LSTM model with technical indicators for intraday trading of gold (XAUUSD) on a 15-minute time horizon. The study examines different combinations of indicators to determine which set produces the most effective trading strategy. The proposed approach is evaluated using backtesting on historical data and forward testing on live market conditions. Performance is assessed using quantitative metrics commonly used in trading problems, such as Win Rate, Sharpe Ratio, Expectancy in Risk Units and Profit Factor. The backtest results demonstrate that the choice of indicators substantially affects strategy performance. The combination of SMA, MACD and Bollinger Bands produced the strongest results, achieving a Sharpe ratio of $2.51$ and a win rate of $38.4\%$, while SMA was identified as the most consistently valuable indicator across configurations. However, forward testing on live market conditions raised concerns about the robustness of these results as none of the strategies achieved positive expectancy and the indicator rankings slightly shifted. Such results suggest that the observed backtest performance was at least partly driven by overfitting to historical market conditions.