A Multi-Source Data Fusion and PSO-Transformer Decision Support System for Gold Price Volatility Forecasting and Risk Warning
4 affiliations
- 1School of Computer Science, Qinghai Normal University, Xining 810008, China
- 2Beijing Key Laboratory of Internet Culture and Digital Dissemination Research, Beijing Information Science and Technology University, Beijing 100192, China
- 3Bank of Qinghai Co., Ltd., Xining 810000, China
- 4School of Computer and Mathematical Sciences, The University of Adelaide, Adelaide, SA 5005, Australia
Abstract
Gold volatility responds to shocks and sentiment, yet the two are rarely fused, and forecasts seldom become actionable. We link a multi-source fusion subsystem to a risk-warning layer through a Transformer whose three pre-registered hyperparameters are searched by particle swarm optimization (PSO-Transformer). It fuses video sentiment, headline scoring and market factors screened to 20 of 21 candidates; the decision side maps the predicted realized variance RVt/Mt (bp2, non-annualized) into three traffic-light tiers of a calibrated predictive distribution. On 5-min spot gold data (January–June 2025, 126 trading days), the PSO-Transformer attains a one-step MAE of 8.038, a 36.5% average improvement over seven baselines under protocol P1, where selection and evaluation share one window; under the leak-free protocol P2, the same pipeline attains 10.16 against 8.82 for the GRU, so the first place reflects the P1 design, not an independent out-of-sample comparison. Diebold–Mariano gains over GARCH, SVM and XGBoost survive a Holm correction. Walk-forward backtesting over 59 origins with persistence-anchored alerts yields 67.8% tier accuracy, 78.8% high-risk precision and 92.9% recall, with the neural engine fed into the same rule attaining 45.5% against 61.4%. One instrument, six months: a documented architecture, not a demonstrated transfer.
Related
- Latent Similarity Gaussian Processes: A Theory-Grounded Approach to Personalized Suicide-Risk Forecasting for Clinical Decision-Support
- A Design Concept of Forecasting Software for Normalized Vector Autoregressions with Fat Tails and Stochastic Volatility
- Forecasting Non-Fungible Token (NFT) Market Volatility Through Hybrid GARCH–LSTM Models
- Volatility-Corrected Prediction Intervals: Locally Adaptive Uncertainty Quantification for Probabilistic Time-Series Forecasting
- Neural Realized Measures: Forecast-Supervised Intraday Aggregation for Volatility Forecasting
BibTeX
@article{diao2026multisource,
title = {{A Multi-Source Data Fusion and PSO-Transformer Decision Support System for Gold Price Volatility Forecasting and Risk Warning}},
author = {Xueping Diao and Xueqiang Lv and Weizheng Wang and Junyi Wu},
journal = {Mathematics},
year = {2026},
month = nov,
doi = {10.3390/math14214115},
url = {https://www.mdpi.com/2227-7390/14/21/4115},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A Multi-Source Data Fusion and PSO-Transformer Decision Support System for Gold Price Volatility Forecasting and Risk Warning
AU - Xueping Diao
AU - Xueqiang Lv
AU - Weizheng Wang
AU - Junyi Wu
PY - 2026
DA - 2026/11/09
JO - Mathematics
DO - 10.3390/math14214115
UR - https://www.mdpi.com/2227-7390/14/21/4115
AB - Gold volatility responds to shocks and sentiment, yet the two are rarely fused, and forecasts seldom become actionable. We link a multi-source fusion subsystem to a risk-warning layer through a Transformer whose three pre-registered hyperparameters are searched by particle swarm optimization (PSO-Transformer). It fuses video sentiment, headline scoring and market factors screened to 20 of 21 candidates; the decision side maps the predicted realized variance RVt/Mt (bp2, non-annualized) into three traffic-light tiers of a calibrated predictive distribution. On 5-min spot gold data (January–June 2025, 126 trading days), the PSO-Transformer attains a one-step MAE of 8.038, a 36.5% average improvement over seven baselines under protocol P1, where selection and evaluation share one window; under the leak-free protocol P2, the same pipeline attains 10.16 against 8.82 for the GRU, so the first place reflects the P1 design, not an independent out-of-sample comparison. Diebold–Mariano gains over GARCH, SVM and XGBoost survive a Holm correction. Walk-forward backtesting over 59 origins with persistence-anchored alerts yields 67.8% tier accuracy, 78.8% high-risk precision and 92.9% recall, with the neural engine fed into the same rule attaining 45.5% against 61.4%. One instrument, six months: a documented architecture, not a demonstrated transfer.
SN - 2227-7390
ER -