Precious metals, particularly gold and silver, have long been valued for their role as safe-haven assets, providing a hedge against currency devaluation and inflation and during periods of economic uncertainty (Smales and Lucey 2019). Physically backed precious metal exchange-traded funds (ETFs) gain exposure through custodied bullion. Authorized participants (APs) create and redeem shares in the primary market to keep prices aligned with net asset value. In times of stress, however, AP balance sheet frictions can widen premiums and discounts and increase tracking slippage. Persistent frictions-expense ratios and secondary-market execution costs-interact with strategy turnover and thus with realized performance. Precious metal ETFs, which provide investors an easy and affordable way to access and invest in these commodities, have grown in popularity due to their liquidity and ability to diversify portfolios.
Accurate predictions of precious metal ETF performance are crucial for financial research as they can enhance portfolio management strategies, optimize asset allocation, and improve risk management practices (Joshi and Dash 2024; Yadav et al. 2024). Hadad et al. (2024) observe that ETFs tend to underperform U.S. equities and provide limited diversification advantages. Nevertheless, during periods of market turbulence, such as the COVID-19 crisis, they proved resilient and reduced downside risk for investors. Other studies corroborate these findings. Pan (2018) finds evidence of explosive behavior in silver and gold prices during the 2008 and 2011 financial crises, suggesting that price bubbles are more likely when the volatility index (VIX) rises. Akhtaruzzaman et al. (2024) highlight greater hedging effectiveness when impact investing is paired with precious metals during crises. Similarly, Rompotis (2024) examines the performance of iShares ETFs in the UK and identifies underperformance linked to replication strategies and the ETF’s maturity.
Forecasting precious metal ETF share prices presents several significant challenges that complicate accurate prediction and reliable investment strategies. One primary issue is the high volatility of precious metal prices, influenced by a myriad of factors including fluctuating investor sentiment (Maghyereh and Abdoh 2022), changing economic conditions (Luo and Ye 2015; Degiannakis and Potamia 2017), and unpredictable geopolitical events (Wang et al. 2019a). Another problem is the noise and bias in sentiment data, which can obscure true market signals and lead to erroneous predictions (Todorov 2024). Additionally, the interconnectedness of global markets means that precious metal ETFs are sensitive to movements in other financial instruments, such as currencies, stocks, and commodities (Todorov 2024). Moreover, traditional statistical forecasting models may struggle to adapt to the rapidly changing market dynamics and the introduction of novel financial products. Finally, high-quality datasets remain scarce. Few sources integrate both quantitative market data and qualitative sentiment information. This limitation makes it difficult to develop robust and accurate predictive models for precious metal ETFs.
A significant knowledge gap in the area of precious metal ETF forecasting is the integration of investor sentiment analysis with traditional quantitative models. Extensive research has examined precious metal price forecasting using macroeconomic indicators and historical price data (Jabeur et al. 2024; Menéndez-García et al. 2024; Shang and Hamori 2024). However, the role of investor sentiment-derived from financial markets, news articles, and online sources-remains underexplored. Bridging this gap requires the development of methods that can process diverse investor sentiment data and effectively integrate it into forecasting frameworks.
Existing AI-based predictive models have improved forecasting accuracy. However, they still face important limitations that reduce their practical effectiveness in dynamic financial markets (Kristjanpoller and Minutolo 2015; Vidal and Kristjanpoller 2020). A key issue is their lack of interpretability and transparency, which reduces investor confidence and limits their use in real-world decision-making contexts (Foroutan and Lahmiri 2024). In addition, the time series of precious metal prices are inherently volatile and influenced by multiple factors, such as geopolitical risk and cross-commodity linkages, making reliable modeling challenging (Mensi et al. 2024). Another shortcoming is that most existing models overlook the influence of investor sentiment on commodity prices, even though sentiment has been shown to significantly affect asset prices in conventional financial markets (Piñeiro-Chousa et al. 2018; Smales 2014).
Existing machine learning (ML)-based trading strategies demonstrate how techniques such as random forests and penalized regression can substantially enhance asset pricing models (Gu et al. 2020). Additionally, Huck (2019) illustrates how large financial datasets and robust machine learning pipelines can be leveraged to execute profitable statistical arbitrage strategies while managing the complexity of big data in financial markets. Turning to deep learning-based trading, Vidal and Kristjanpoller (2020) show that a CNN-LSTM hybrid can yield high-accuracy volatility predictions, particularly in commodity markets such as gold. Meanwhile, reinforcement learning approaches have gained traction due to their adaptive decision-making capabilities. Deng et al. (2017) present a deep direct reinforcement learning method that actively learns trading signals from raw price data. Similarly, Huang et al. (2024) introduce a BiLSTM attention-based deep reinforcement learning framework for portfolio management, illustrating how attention mechanisms help the agent focus on critical market features when making trading decisions.
However, machine learning-based trading systems typically focus on predicting prices rather than optimizing trading decisions, and they do not adapt in real-time to sudden market shifts. Deep learning methods, while powerful, often struggle in volatile or low-liquidity environments where training data are sparse or noisy. Reinforcement learning methods tend to be more flexible in adapting to changes such as volatility spikes or regime shifts, but they can suffer from stability issues-particularly when markets change abruptly.
Crucially, current forecasting models and trading strategies tend to operate in a static or rule-based manner, lacking the ability to respond adaptively to rapid market changes. Traditional approaches such as buy-and-hold or technical trading rules struggle with market inefficiencies and cannot dynamically adapt to changing conditions. While machine learning models improve price forecasting, they often fail to translate forecasts into optimized trading decisions. Algorithmic strategies, on the other hand, tend to focus on execution speed rather than strategic adaptability or profitability. This highlights a significant gap in the literature: the lack of intelligent, adaptive trading systems that combine predictive modeling with real-time decision-making. Filling this gap requires the integration of forecasting tools with reinforcement learning techniques capable of optimizing long-term returns in volatile and uncertain environments.
Our proposed hybrid SVR-PPO approach aims to address these gaps by:
* Including investor sentiment factors, thereby capturing shifts in market mood or fear.
* Combining interpretable price forecasts using support vector regression (SVR) with an adaptive reinforcement learning component, which steps in when pure forecasting signals prove inadequate-especially in volatile phases.
* Offering a transparent rule-based “switch” that integrates sentiment-informed prediction with reinforcement learning (RL)-driven trade execution, thereby balancing accuracy, adaptability, and interpretability.
Overall, the key contributions of this study are threefold:
* The study proposes a novel hybrid trading system that combines SVR for price forecasting with a PPO (proximal policy optimization) reinforcement learning model for adaptive decision making. This integration enables both interpretable forecasts and dynamic, optimized trading actions, thereby improving performance under changing market conditions. This dual-system approach exploits the interpretability and accuracy of SVR-based predictions while dynamically adapting trading decisions using reinforcement learning, especially during periods of SVR underperformance. Empirical results show that this hybrid SVR-PPO strategy significantly outperforms traditional benchmarks – such as buy-and-hold and moving average strategies – not only in terms of cumulative returns, but also across a range of risk-adjusted performance measures
* This study highlights the predictive value of market-based VIX (Chicago Board Options Exchange’s CBOE Volatility Index) and news-based SSW (Shapiro, Sudhof, Wilson) investor sentiment indices when integrated with financial data. Here we show that the inclusion of these sentiment indicators significantly improves forecasting accuracy and directional prediction for precious metals ETFs.
* Empirical evaluation shows that the proposed hybrid strategy consistently outperforms baseline methods-such as the 5-day moving average and buy-and-hold-in terms of cumulative returns, Sharpe ratio and trading profitability across all four targeted ETFs, demonstrating its robustness and effectiveness.
The rest of the paper is organized as follows. Section 2 presents the theoretical background and develops the research hypotheses. Section 3 reviews the relevant literature on commodity ETFs, forecasting techniques, investor sentiment and trading strategies. Section 4 describes the dataset and input variables. Section 5 presents the construction of the hybrid SVR-PPO trading framework. Section 6 presents the experimental setup, model configurations, and performance evaluation metrics, followed by empirical results comparing our approach with benchmark models. This section also examines the robustness of the proposed strategy under different trading scenarios. The results and their practical implications are discussed in Sect. 7. Finally, Sect. 8 concludes the paper with key findings, limitations and suggestions for future research.
