Predicting Precious Metal Markets Using Custom Deep-Learning NARX Network: A Portfolio Application

Precious metals such as gold, silver, platinum, and palladium represent beneficial investments due to their haven characteristics. Including these metals in portfolios can avoid substantial losses, especially during a financial crisis. Constructing and implementing efficient portfolios for financial markets requires advanced statistical data analysis, including studying external impact, predictive analysis, and risk-return investigation. This chapter presents the construction and examination of diversified precious metal portfolios representing a three-month (July - eptember 2021) investment during the coronavirus pandemic economic downturn. A total of 58 assets from global futures, equity, and Exchange-Traded Fund (ETF) markets were investigated as potential investment targets by conducting preliminary research to identify factors affecting asset prices. The correlations among these factors were examined using historical data from July 2016 to June 2021 (5-year period) using an Elastic Net (EN) feature selection method. The dataset was filtered by removing features representing low correlations with the selected assets. Relationships among 17 assets could not be confirmed due to low performance of EN, leading to speculation regarding the widely discussed haven behavior of precious metals. The updated dataset from EN was utilized for training a deep learning neural network called Nonlinear AutoRegressive with eXogenous input (NARX), which has proved to be effective for time series predictions. The performance of NARX model was optimized via a One-Factor-at-a-Time (OFAT) approach by utilizing step changes in a training parameter called Learning Rate (LR). The optimum configuration achieved was implemented to generate multi-step price predictions over the selected period. Forecast errors from each asset were compared through average percentage error and maximum deviations, where the total sum should be less than 10% for assets to exhibit high performance. Further optimization of NARX performance was suggested by employing multi-objective approaches requiring a higher degree of complexity. Based on their prediction error, a portfolio set representing nine assets was chosen for further examination to measure risk-return levels. Portfolio returns were optimized through Mean-Variance (MV) and Conditional-Value-atRisk (CVaR) strategies by comparing their results with actual data. Both strategies produced positive returns, although predicted MV outperformed CVaR by a minor margin. The predicted CVaR did not include three target assets in the actual portfolio due to relatively high error deviations. A refined maximization of portfolio returns was suggested by improving prediction performance and modifying target returns. The findings suggest that portfolio performance is susceptible to prediction accuracy, implicating that neural network optimization is crucial. This study contributes towards data analytics within precious metal markets using deep learning artificial intelligence. Furthermore, it reveals valuable information regarding the behavior of these markets during a recession. These revelations allow investors to develop reliable diversification strategies to safeguard their assets amid a plunge in the market.