Design and Evaluation of an Advanced Attention-Based Deep Learning Framework for Accurate, Robust, and Interpretable Forecasting of Global Gold Prices with a Focus on the Iranian Market
Keywords:
Gold Price Forecasting, Temporal Fusion Transformer, Deep Learning, Attention Mechanism, Multi-Horizon Forecasting, Interpretability, Probabilistic Forecasting, Iranian MarketAbstract
This study aimed to design and evaluate an advanced Temporal Fusion Transformer-based deep learning framework for accurate, robust, probabilistic, multi-horizon, and interpretable forecasting of global gold prices with particular relevance to the Iranian market. A quantitative time-series forecasting design was applied to daily gold-market data covering January 19, 2014, to January 22, 2024. The dataset included Open, High, Low, Close, and Volume observations and was divided chronologically into training, validation, and test subsets. The proposed Temporal Fusion Transformer integrated recurrent temporal processing, variable selection networks, gated residual networks, bidirectional LSTM encoding, interpretable multi-head attention, and quantile regression. Forecasts were generated across multiple horizons and evaluated using RMSE, MAE, MAPE, , Prediction Interval Coverage Probability, Prediction Interval Normalized Range Width, and Mean Directional Accuracy. Comparative benchmarks included ARIMA, SARIMA, LSTM, GRU, standard Transformer, and Informer. Statistical differences in predictive accuracy were assessed using the Diebold-Mariano test. The proposed TFT significantly outperformed ARIMA ( ), SARIMA ( ), LSTM ( ), GRU ( ), and the standard Transformer ( ). Its superiority over Informer was not statistically significant at the 95% confidence level ( ), although significance was achieved at the 90% level. The model also preserved its comparative advantage as the forecasting horizon increased, indicating greater robustness in longer-horizon prediction. The proposed TFT framework provides a statistically supported and practically useful approach for global gold-price forecasting by combining strong predictive performance, robustness across horizons, calibrated uncertainty estimation, and intrinsic interpretability, thereby offering potential value for investment analysis and economic decision-making in the Iranian market.
References
[1] Y. Guo, C. Li, X. Wang, and Y. Duan, "Gold Price Prediction Using Two-Layer Decomposition and XGBoost Optimized by the Whale Optimization Algorithm," Computational Economics, vol. 66, no. 2, pp. 1157-1189, 2025, doi: 10.1007/s10614-024-10736-9.
[2] P. D. Quang, "Analysis and Forecasting of Daily Global Gold Price: An SARIMA-LSTM Approach with Random Forest Technique," Cogent Economics & Finance, vol. 13, no. 1, p. 2568969, 2025, doi: 10.1080/23322039.2025.2568969.
[3] Y. Wang and T. Lin, "A Novel Deterministic Probabilistic Forecasting Framework for Gold Price with a New Pandemic Index Based on Quantile Regression Deep Learning and Multi-Objective Optimization," Mathematics, vol. 12, no. 1, p. 29, 2024, doi: 10.3390/math12010029.
[4] A. Amini and R. Kalantari, "Gold Price Prediction by a CNN-Bi-LSTM Model along with Automatic Parameter Tuning," PLOS ONE, vol. 19, no. 3, p. e0298426, 2024, doi: 10.1371/journal.pone.0298426.
[5] Y. Lin, "An Improved CNN-LSTM Model for Daily Gold Price Prediction," 2026: Atlantis Press, pp. 608-621, doi: 10.2991/978-94-6239-648-7_53.
[6] J. Wu and J. Liu, "Gold Price Prediction Based on LSTM-Attention Combined Model," Open Access Library Journal, vol. 13, p. e1114759, 2026, doi: 10.4236/oalib.1114759.
[7] S. Liu and Q. Lan, "Long-Term Gold Price Prediction Using the CNN-BiGRU Model Enhanced with an Attention Mechanism," 2025: IEEE, pp. 1369-1375, doi: 10.1109/ICAACE65325.2025.11020201.
[8] B. A. Memon, R. Tahir, H. M. Naveed, and K. Cheng, "Forecasting Gold and Platinum Prices with an Enhanced GRU Model Using Multi-Headed Attention and Skip Connection," Mineral Economics, vol. 38, no. 4, 2025, doi: 10.1007/s13563-025-00520-y.
[9] Y. Yang, "TCN-QV: An Attention-Based Deep Learning Method for Long Sequence Time-Series Forecasting of Gold Prices," PLOS ONE, vol. 20, no. 5, p. e0319776, 2025, doi: 10.1371/journal.pone.0319776.
[10] Y. Zhao, Y. Guo, and X. Wang, "Hybrid LSTM-Transformer Architecture with Multi-Scale Feature Fusion for High-Accuracy Gold Futures Price Forecasting," Mathematics, vol. 13, no. 10, p. 1551, 2025, doi: 10.3390/math13101551.
[11] C. Izzo, "On Explainable Deep Learning for Macroeconomic Forecasting and Finance," UCL (University College London), 2022.
[12] B. Lim, S. Ö. Arık, N. Loeff, and T. Pfister, "Temporal Fusion Transformers for Interpretable Multi-Horizon Time Series Forecasting," International Journal of Forecasting, vol. 37, no. 4, pp. 1748-1764, 2021, doi: 10.1016/j.ijforecast.2021.03.012.
[13] V. Egy, "PERBANDINGAN KINERJA MODEL TFT DAN N-BEATS DALAM PREDIKSI HARGA EMAS DENGAN INKORPORASI FITUR MAKROEKONOMI," ed, 2025.
[14] W. Ge, P. Lalbakhsh, L. Isai, A. Lensky, and H. Suominen, "Comparing Deep Learning Models for the Task of Volatility Prediction Using Multivariate Data," 2023.
Downloads
Published
Issue
Section
License
Copyright (c) 2025 Morteza Mohammadlu, Seyed Mohsen Mirhosseini (Author)
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
