Comparative Evaluation of Deep Learning Architectures for Gold Price Forecasting: a Case Study of Hierarchical vs. Transformer-Based Approaches

Abstract

Forecasting commodity prices, particularly gold, is a critical challenge in financial research due to the market's dynamic and volatile nature. Gold serves as a safe-haven asset, yet its price trajectory is influenced by a complex interplay of macroeconomic indicators, geopolitical tensions, and market sentiment, making accurate prediction a formidable task. This study evaluates the predictive capabilities of two contrasting deep learning architectures—Chronos (a hierarchical temporal model) and PatchTST (a patch-based Transformer)—to assess their effectiveness in forecasting gold prices. The research utilizes historical daily closing prices from August 2000 to November 2024, enriched with 24 technical indicators including momentum, volatility, and trend measures. The models were trained on a 90-day lookback window to forecast a 10-day horizon using a rigorous sliding window approach. Results indicate a stark divergence in performance: Chronos achieved excellent accuracy with a Mean Absolute Percentage Error (MAPE) of 4.867% and an R² of 79.75%, demonstrating the effectiveness of hierarchical modeling for trend-persistent commodities. Conversely, PatchTST failed catastrophically with a MAPE of 44.485% and a negative R² of -1071.33%, highlighting the fundamental risks of applying channel-independent transformer architectures to univariate financial data where feature correlation is essential. The findings provide empirical evidence that hierarchical architectures are superior for gold price forecasting, while rigid transformer adaptations require careful validation against market microstructure.