Forecasting Metal Futures Using Machine Learning and Deep Learning

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.May 10 2024

In a paper published in the journal Resources Policy, researchers explored forecasting metal futures in commodity markets using machine learning (ML) and deep learning (DL) models, assessing their performance with metrics like root mean squared error (RMSE) and mean absolute percentage error (MAPE).

The study stood out for its comprehensive approach, analyzing multiple metal commodity futures simultaneously and incorporating both ML and DL techniques. Despite promising results, the research highlighted the influence of factors like metal choice and sample period on prediction performance, cautioning against simplistic theoretical conclusions solely based on these models.

Related Work

Previous research has predominantly relied on statistical techniques such as autoregression and moving averages for forecasting commodity prices. Yet, due to computational constraints and unrealistic assumptions, interest has shifted towards ML algorithms. Models like artificial neural networks (ANN) and support vector regressor (SVR) have demonstrated effectiveness in capturing the dynamic nature of commodity markets, outperforming traditional statistical methods across various financial asset classes.

Deep learning (DL) models, notably long short-term memory (LSTM) and convolutional neural networks (CNN) have emerged as powerful tools for predicting metal prices, surpassing conventional ML approaches. While recent studies have explored hybrid models, their application in predicting precious metal commodities warrants further investigation.

Metal Price Forecasting

The study collected daily future prices of metals, covering gold, silver, copper, platinum, palladium, and aluminum, from February 2003 to January 2023, obtained from Yahoo Finance. Pre-processing steps were implemented to handle missing values and scale the data using Min-Max scaling. This scaling technique was crucial for the efficient convergence of ML models. Subsequently, the data transformed input-output pairs suitable for DL models, facilitating the exploration of various ML and DL methodologies.

Among the methodologies employed, significant attention was directed toward the LSTM model due to its proficiency in capturing long-term dependencies within time series data. Stemming from recurrent neural networks (RNNs), LSTM models are adept at modeling non-linear relationships and overcoming the vanishing gradient problem inherent in traditional RNNs. Moreover, the study delved into advanced DL models such as Stacked-LSTM, Gated Recurrent Unit (GRU), and bidirectional Long-Short Term Memory (Bi-LSTM) to enhance forecasting accuracy.

The study integrated the SVR and extreme gradient boosting (XGBoost) algorithms, recognized for their effectiveness in addressing non-linear relationships and temporal dependencies prevalent in time series forecasting. The methodology involved transparent parameter settings for all models, ensuring reproducibility and facilitating comprehensive understanding. Additionally, the study conducted robustness checks by comparing results obtained using both 60-day and 30-day input periods, contributing to the reliability of the findings.

Metal Price Forecasting

Various metrics, such as RMSE, MAE, and MAPE, were employed to evaluate the performance of ML and DL models in forecasting metal prices. RMSE quantifies the squared deviation between observed and predicted values, emphasizing accuracy by penalizing large errors.

Meanwhile, MAE provides a straightforward measure of discrepancy between observed and predicted values without squaring errors, offering robustness to outliers. Lastly, MAPE evaluates accuracy by measuring the average percentage difference between forecasted and actual values, providing insight into how well predictions align with actual data.

The comparative analysis of different models revealed variations in performance across metals and data samples. Models like stacked-LSTM, bidirectional LSTM, convolution LSTM, and GRU demonstrated superior forecasting capabilities for certain metals, while SVR exhibited lower performance for others. Sub-sample analysis further highlighted the influence of data period on model performance, indicating no consensus on the best-performing model across all scenarios. Moreover, models' performance varied with input periods, suggesting the importance of considering recent data for improved predictions.

Despite the challenges of market dynamics and inherent uncertainties, ML and DL models showcased promising capabilities in forecasting metal prices. However, the study emphasized the need for continuous innovation to enhance model accuracy and robustness.

Market efficiency, varying asset characteristics, and external factors like geopolitical events further contribute to the complexity of price prediction. Nonetheless, these findings have significant implications for portfolio management, hedging, and risk mitigation strategies, underscoring the importance of informed decision-making in financial markets.

Conclusion

To sum up, this study delved into forecasting metal futures using various ML and DL techniques, including stacked-LSTM, convolutional LSTM, SVR, bidirectional LSTM, XGBoost, and GRU. Through analysis spanning 20 years of data and sub-sample assessments, it became evident that model performance varied across metals, input periods, and time frames, with no consistently superior model identified.

The findings align with the efficient market hypothesis, suggesting challenges in consistently outperforming the market in predicting metal prices. Diversification of investment strategies is recommended due to the heterogeneity in predictive performance. Future research should explore hybrid models integrating traditional ML and DL approaches and consider deep reinforcement learning for adaptive trading strategies in commodity markets.