Multi-Timeframe Analysis for XAU/USD Algorithmic Trading: A Comparative Study of PatchTST, Deep Learning, and Ensemble Methods

The global financial market is extremely volatile, especially for the safe-haven asset XAU/USD (Gold), which sets new All-Time Highs (ATHs) consistently. The noisy Over-theCounter (OTC) market data and the extrapolation errors during the ATH conditions pose challenges to older machine learning models and traditional forecasting methods, compounded by the absence of a unified multi-timeframe prediction context in prior work. Which proposes comparative algorithmic forecasting engine according to the Cross-Industry Standard Process for Data Mining (CRISP-DM). The study evaluates the performance of the Patch Time Series Transformer (PatchTST), Deep Learning (LSTM), and Ensemble Methods (XGBoost, Random Forest) at different time scales. The models use 3000 hours of the historical high-frequency OHLC (Open, High, Low, Close) data of XAU/USD, and the last 150 hours are used for testing. To reduce extrapolation bias, a 'Delta Transformed' median price system is used. The results show that PatchTST achieves a significant increase in the ability to detect extreme wicks for the 'Sniper Entries' while LSTM works well in predicting a sequence of continuous data. Furthermore, Random Forest is capable of filtering the ultra-high-frequency microstructure noise. Finally, this research provides a predictive model to connect raw computation with real market analysis in very dynamic Over-theCounter markets