Resources Policy 92 (2024) 105040

Contents lists available at ScienceDirect

Resources Policy
journal homepage: www.elsevier.com/locate/resourpol

How good are different machine and deep learning models in forecasting
the future price of metals? Full sample versus sub-sample
Anu Varshini a, Parthajit Kayal a, Moinak Maiti b, 1, *
a
Madras School of Economics (MSE), Gandhi Mandapam Road, Behind Government Data Centre, Kottur, Chennai, 600025, India
b
Department of Finance, School of Economics and Finance, University of the Witwatersrand, Johannesburg, South Africa

A R T I C L E I N F O A B S T R A C T

Keywords: This study aims to forecast metal futures in commodity markets, including gold, silver, copper, platinum,
Machine learning palladium, and aluminium, using different machine and deep learning models. Prevalent models such as Stacked
Deep learning Long-Short Term Memory, Convolutional LSTM, Bidirectional LSTM, Support Vector Regressor, Extreme
Forecasting
Gradient Boosting, and Gated Recurrent Unit are utilized. The model performance is assessed by multiple factors
Commodity
Futures
such as Root Mean Squared Error, Mean Absolute Error, and Mean Absolute Percentage Error. The study stands
out by considering multiple metal commodity futures simultaneously, incorporating both Machine Learning and
Deep Learning models, and conducting two sets of experiments with a full sample and subsample analysis. In
addition, it uses different inputs of 30- and 60-days periods for robustness checks. Mean Absolute Percentage
Error values suggest that different machine and deep learning models are efficient on prediction the future metal
prices. However, the model performance varies significantly with the influence of metal choice, sample period,
and inputs on prediction performance. Therefore, it suggests that constructing a theory based on machine and
deep learning models is challenging.

1. Introduction dynamics on commodity prices (Fama, 1984; Verheyen, 2010). Com­

modity futures markets have played a vital role in reducing price vola­
Commodity markets play a significant role in the global economy, tility, enhancing market efficiency, and providing benefits to producers
offering benefits like as price discovery, hedging opportunities, and and consumers by mitigating the risk of possible price shocks (Jacks
ensuring liquidity. They are particularly instrumental in propelling et al., 2011).
economic growth, especially in resource-rich nations (Chang and Fang, Recent studies underscore the significance of employing artificial
2022). Extensive research has underscored the importance of accurate intelligence (AI) techniques, such as machine learning (ML) and deep
price predictions in commodity markets, offering valuable insights for learning (DL) algorithms, to forecast commodity prices. These tech­
traders, investors, and policymakers. By assessing the price volatility, niques have demonstrated substantial improvements, achieving accu­
traders and investors can hedge their positions and mitigate potential racy and robustness improvements of over 60% (Li et al., 2017).
losses. In a recent study, Kumar et al. (2022) emphasize that price pre­ Previous work primarily relied on traditional time series econometric
diction can help manage risks, optimize profits, and enhance market techniques like ARIMA, ARCH, and GARCH models (Arowolo, 2013;
efficiency. Moreover, reliable price forecasts are important for Bhardwaj et al., 2014). However, it has been demonstrated that these
improving portfolio performance and guiding investment decisions models struggle to capture non-linear patterns, prompting the adoption
(Horák and Jannová, 2023). Investing in commodity futures helps in­ of ML and Artificial Neural Networks (ANN) as alternatives, particularly
vestors achieve excess returns. Further, changes in commodity prices in the presence of non-linearity (Shahwan and Odening, 2007). ML
reflects direction of future inflation. This relationship is particularly methodologies proficiently identify stochastic variables and capture
noticeable for future prices compared to spot prices (Fama, 1984). Un­ latent non-linear characteristics that conventional econometric models
like spot prices, future prices provide more accurate forecasts that are incapable of detecting (Herrera et al., 2019). Artificial Recurrent
consider inflation and effectively reflect the impact of supply-demand Neural Networks (ARNN) equipped with memory, such as Long-Short

* Corresponding author.
E-mail addresses: [email protected], [email protected] (M. Maiti).
1
Moinak Maiti was Independent Researcher, India, when this paper was submitted to the journal.

https://doi.org/10.1016/j.resourpol.2024.105040
Received 24 August 2023; Received in revised form 3 January 2024; Accepted 25 April 2024
Available online 30 April 2024
0301-4207/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-
nc/4.0/).
A. Varshini et al. Resources Policy 92 (2024) 105040

Term Memory (LSTM), have demonstrated considerable superiority over distribution (Parmezan et al., 2019).
ARIMA models in terms of performance (Siami-Namini et al., 2019). The emergence of ML techniques has prompted a shift in the focus of
Furthermore, hybrid models have gained popularity due to their commodity price forecasting from statistical methods to ML-based ap­
improved forecasting accuracy, though conflicting results persist proaches. ML algorithms bring the advantage of adeptly managing
regarding their superiority (Herrera et al., 2019). extensive and complex datasets prevalent in commodity markets, which
Metals represent a substantial share of commodity markets and as­ are characterized by substantial volatility and uncertainty stemming
sume a pivotal role in the global economy, being influenced by diverse from geopolitical events, macroeconomic trends, and supply chain dy­
economic and geopolitical factors (Gong and Xu, 2022). Given the namics (Verheyen, 2010). These algorithms are aptly equipped to cap­
growing interconnectedness of factors influencing metal prices, accurate ture the dynamic nature of commodity markets, where prices experience
predictions of future metal prices present challenges for investors and frequent fluctuations (Schnepf et al., 2005). Time series forecasting,
speculators seeking risk management. The primary aim of this research with the availability of abundant data and increased computing power,
is to employ various machine learning (ML), and deep learning (DL) has become an area of extensive research, and ML models have become
techniques to forecast metal futures within commodity markets. The an integral part of it (Lim and Zohren, 2021). Unlike traditional statis­
metals under consideration encompass gold, silver, copper, platinum, tical methods, ML algorithms excel at identifying patterns in sequential
palladium, and aluminium. The analysis focuses on well-established data that exhibit time-dependent patterns, which can be challenging for
models in time series analysis, namely Stacked-LSTM (Li et al., 2020; traditional methods to analyze (Herrera et al., 2019). As an example,
Ojo et al., 2019), Convolutional LSTM (Essien and Giannetti, 2019), Artificial Neural Networks (ANN) have demonstrated superior perfor­
Bidirectional LSTM (Siami-Namini et al., 2019), Support Vector Re­ mance compared to classical statistical methods such as linear regres­
gressor (SVR) Model (Makala and Li, 2021), Extreme Gradient Boosting sion and Box-Jenkins approaches when forecasting excess returns in
(XGBoost) model (Jabeur et al., 2021), and Gated Recurrent Unit (GRU) large stocks (Desai and Bharati, 1998). Similar findings are reported by
model (Torres and Qiu, 2018). The performance of these models is Krollner et al. (2010) in their study on modelling non-linear time series
evaluated by Root Mean Squared Error (RMSE), Mean Absolute Error using artificial neural networks. Other popular ML models used in time
(MAE), and Mean Absolute Percentage Error (MAPE) metrics. Although series forecasting include Support Vector Regressor (SVR) (Cao and Tay,
some researchers favour MAE over RMSE, it is advantageous to consider 2003), Extreme Gradient Boosting (XGBoost) (Zhang et al., 2021b), and
a combination of these indicators (Chai and Draxler, 2014). Recurrent Neural Network (RNN) such as Long-Short Term Memory
This study differentiates itself from previous literature in several (LSTM) and Gated Recurrent Unit (GRU), which capture temporal de­
aspects. First, it considers multiple metal commodity futures simulta­ pendencies and patterns in data (Yamak et al., 2019). Hybrid models,
neously, which is unique compared to previous work focused on pre­ including CNN-LSTM, LSTM-CNN, ConvLSTM, and Encoder-Decoder,
dicting the prices of individual metal commodities (Astudillo et al., have exhibited encouraging performance when compared to classical
2020; Hussein et al., 2011; Livieris et al., 2020). Second, it incorporates models (Kilinc and Yurtsever, 2022).
both ML and DL models to derive the study conclusions. Third, the study ML models are increasingly applied in predicting various financial
comprises two sets of experiments: one involving the construction of assets, including stocks, bonds, commodities, and cryptocurrencies, to
prediction models using a full 20-year sample from 2003 to 2023, and assist investors in making informed decisions. In stock markets, ML
the other conducting sub-sample analysis to identify potential models models demonstrate superior performance compared to statistical and
that consistently outperform the rest. Fourth, different robustness econometric models by capturing the non-linear nature of equity mar­
measures are used to validate the study findings such as different input kets (Ampomah et al., 2020). In bond markets, ML algorithms are used
periods. The results of this study do not yield conclusive evidence to predict changes in bond prices and forecast future interest rate
regarding the superiority of any specific model for accurate commodity movements (Bianchi et al., 2021). In commodity markets and crypto­
price forecasting. The choice of commodities, sample periods, and inputs currencies, ML models help identify hidden patterns and forecast price
significantly influences the predictive performance of the models, sug­ movements. Notable models used in these asset classes include.
gesting the absence of a consistently superior model. The study findings
regarding the variations in model performance across metals are highly • Stocks - SVM (Reddy, 2018), Random Forest (RF), LSTM (Nikou
relevant in the context of the Efficient Market Hypothesis, Portfolio et al., 2019); ANN (Maiti et al., 2022; Vukovic et al., 2020)
Theory, Behavioural finance, and hedging strategies. • Bonds – Linear Regression, Boosted Regression Trees, RF, and Deep
section 2 offers an extensive literature review, drawing comparisons Neural Networks (Bianchi et al., 2021)
between our research and previous studies. Section 3 describes the data • Commodity Markets – ARIMA, SVR, Prophet, XGBoost and LSTM
source and the models employed in this study to predict future com­ (Chen et al., 2021)
modity prices. The analysis, results, and subsequent discussion are • Cryptocurrencies – Neural Networks, SVR, RF (Valencia et al., 2019);
offered in Section 4. Finally, the study concludes in the last section, ANN, LSTM (Maiti et al., 2020), XGBoost (Ranjan et al., 2023)
providing recommendations for future research.
The prices of metals in commodity markets exhibit extreme volatility
2. Review of literature and are challenging to predict, necessitating appropriate modelling
techniques (Kayal and Maheswaran, 2021). Deep learning (DL) models
For more than five decades, autoregression (AR) and moving aver­ have exhibited substantial advancements in the literature of financial
ages (MA), which are statistical techniques, have been considered the time series forecasting, surpassing traditional ML models (Jabeur et al.,
cutting edge in time series modelling and prediction (Parmezan et al., 2021). DL approaches excel in extracting key features from complex
2019). Prior studies on forecasting prices in commodity markets have data, and recent advancements in attention mechanisms have further
mainly focused on linear statistical methods, particularly ARIMA enhanced prediction performance beyond baseline models (Zhou et al.,
models. However, these techniques possess limitations concerning their 2019). The study by Zhang et al. (2021a) employ a range of algorithms,
modelling capabilities and computational demands (Zulauf et al., 1999). including Multi-Layer Perceptron Neural Network, SVM, Gradient
Additionally, statistical models often hinge on unrealistic assumptions Boosting Tree, and RF, to forecast monthly copper prices. Similarly, the
like variable independence and normal distribution, which do not align study by Baser et al. (2023) employ robust tree-based prediction models
with real-world scenarios (Onour and Sergi, 2011). In recent times, there such as Decision Tree, Adaptive Boosting, Random Forest (RF), Gradient
has been a surge in interest towards non-parametric modelling using Boosting, and XGBoost to predict the price of Gold, considering the
Machine Learning (ML) algorithms for time series prediction. This presence of outliers commonly observed in commodity markets. RNN,
approach offers enhanced flexibility as it avoids assuming a specific data particularly variations like LSTM and Convolutional Neural Network

2
A. Varshini et al. Resources Policy 92 (2024) 105040

(CNN), are commonly used DL models for commodity price forecasting (RNNs) proposed by Hochreiter and Schmidhuber (1997), enables the
due to their ability to handle complex sequential data in commodity capture of long-term dependencies within time series data. This model
markets (Yamak et al., 2019). Vidya and Hari (2020) used variants of excels at modelling non-linear relationships between input and output
RNN such as CNN and LSTM to forecast gold price series. It is further sequences, rendering it suitable for tackling intricate time-series fore­
extended to syndicate models such as CNN-LSTM, Bi-directional-LSTM, casting challenges. The primary purpose behind developing LSTM was
and Bi-LSTM-CNN models by Peng (2021). However, the utilization of to tackle the vanishing gradient problem encountered by RNNs, which
hybrid approaches in forecasting precious metal commodities is rela­ hinders their ability to handle long-term dependencies (Sherstinsky,
tively novel and has received limited exploration. 2020).
The main objective of this paper is to enhance the existing literature The LSTM architecture, depicted in Fig. 2, incorporates a cell state in
by establishing a connection between metal commodity futures price addition to hidden states, thereby providing the model with an extended
point forecasting and the utilization of various ML and DL techniques. memory of past events. The cell state can be modified by employing
Specifically, the study incorporates Stacked-LSTM, ConvLSTM, Bi- gates, with three types of gates existing in an LSTM model: the forgetting
LSTM, SVR, XGBoost, and GRU models in this endeavour. These gate, input gate, and output gate. The initial step involves determining
models are effective in analyzing time series data with long-term de­ whether to retain or discard information from the previous timestamp.
pendencies, allowing for the extraction of sequential patterns in the This process considers the input of ht− 1 and xt which subsequently passes
data. Additionally, this study considers all available metal futures for through the sigmoid layer, generating values ranging between 0 and 1. If
trading over a 20-year period from 2003 to 2023. Furthermore, we the output is close to 0, the input is considered irrelevant and is not
conduct sub-sample analysis to identify the best model for commodity propagated further. Conversely, if the output is close to 1, the input is
price prediction by detecting structural breaks in the data. deemed relevant and continues to propagate through the network.
( )
Ft = σ Wf ∗ (ht− 1 , xt ) + bf (1)
3. Data and methodology
The subsequent stage involves determining the relevant information
3.1. Data to be incorporated into the cell state, considering the previous hidden
state and input data. This process occurs in two parts (See equations (1)–
The daily future prices of metals (gold, silver, copper, platinum, (6)). Firstly, the input gate, comprising a sigmoid layer, determines
palladium, and aluminium2) from February 2003 to January 2023 were which values should be updated. The resulting output is then multiplied
collected from the Yahoo Finance website. The data was directly im­ by the output from the tanh layer, generating a vector of new candidate
ported into a Jupyter notebook using the yfinance package and stored as values represented as C ̃ t . While the sigmoid function ensures that the
a data frame. Pre-processing steps were applied to handle missing model remains differentiable, the tanh activation function distributes
values, and the data were scaled using Min-Max scaling. The applied the gradients as it has a zero cantered range of − 1 to 1 thereby miti­
scaling technique encompassed subtracting the minimum value of each gating the exploding or vanishing gradient problem and allowing the
feature and subsequently dividing it by its range. Scaling plays a pivotal cell information to flow longer.
role in the construction of machine learning (ML) models due to their
sensitivity to the scale of input features, and it facilitates the more It = σ(Wi ∗ (ht− 1 , xt ) + bi ) (2)
efficient convergence of the gradient descent algorithm. Subsequently,
the data was transformed into a series of input-output pairs suitable for ̃ t = tanh (Wc ∗ (ht− 1 , xt ) + bc )
C (3)
feeding into the deep learning (DL) models. The data is then split into
Subsequently, the old state is multiplied by Ft and add It ∗ C
̃ t . This
training and testing datasets and reshaped to a compatible format for
computation yields the new candidate value, which is appropriately
input into the DL models. The primary analysis includes a window of 60
scaled based on the decision to update each state value.
days’ time periods is chosen as inputs (See Tables 1–3). For robustness
checks the study also runs all the models with 30 days’ time periods to Ct = Ft ∗ Ct− 1
̃t
+ It ∗ C (4)
compare the results (See Tables 4–6).
ML and DL methods have demonstrated effectiveness in addressing In the final output layer, the first step involves applying a sigmoid
time series forecasting problems with complex non-linear relationships function to determine the inclusion of specific components of the cell
among multiple input variables (Shahwan and Odening, 2007). Pa­ state in the output. Following this, the tanh function is applied, resulting
rameters settings of all the selected models are presented in Appendix-I. in the generation of a new hidden state.
Fig. 1 illustrates the normalized closing prices of gold, silver, copper, Ot = σ (Wo ∗ (ht− 1 , xt ) + bo ) (5)
platinum, and palladium. By examining the price series of different
metals in Fig. 1, we observe that palladium exhibits the highest vola­ Ht = Ot ∗ tanh Ct (6)
tility, while platinum has the least. Gold, silver, and copper show similar
levels of volatility, with copper slightly higher than the other two It is Furthermore, LSTM models find applications in various domains,
noteworthy that the volatility levels can vary across different periods, such as machine translation (Zhou et al., 2016), speech recognition (Han
even for the same metal. In financial time series predictions, the level of et al., 2017), video tagging (Gao et al., 2017), and the forecasting of
volatility can significantly impact prediction accuracy. High volatility co-movements in cryptocurrencies (Maiti et al., 2020; Ranjan et al.,
leads to larger price swings and increased uncertainty, making it chal­ 2023).
lenging for models to identify underlying patterns. This distinct vola­
tility patterns among the metals can also be linked to the patterns related 3.2.2. Stacked-LSTM (Stacked Long-Short Term Memory)
to storage costs or supply disruptions (Raza et al., 2023). The Stacked-LSTM model, depicted in Fig. 3, differs from the original
LSTM model in that it has multiple hidden LSTM layers, each containing
multiple memory cells. Due to its increased complexity and depth
3.2. Methodologies
compared to the original model, this advanced model enables us to
achieve even more accurate forecasts. In this architecture, the output of
3.2.1. LSTM model (long -short term memory)
one LSTM layer serves as the input for the subsequent LSTM layer. Each
The LSTM model, which is a variation of Recurrent Neural Network
LSTM layer possesses its own distinct set of parameters, encompassing
the three gates: the input gate, forget gate, and output gate. These gates
2 play a crucial role in regulating the information flow within the layer.
Aluminium data is only available from 2013 onwards.

3
A. Varshini et al. Resources Policy 92 (2024) 105040

Table 1
RMSE, MAE, and MAPE Scores for different models across metals for full sample data for 60-days’ input.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Stacked LTSM Convolutional-LSTM SVR Bidirectional LSTM XG Boost GRU

Gold RMSE 17.85 30.28 21.54 23.85 19.86 18.08

MAE 13.86 25.91 16.79 19.65 15.52 14.16
MAPE 0.74% 0.81% 0.93% 1.01% 0.86% 0.79%
Silver RMSE 0.90 0.49 0.48 0.44 0.51 0.54
MAE 0.77 0.37 0.38 0.34 0.39 0.42
MAPE 3.49% 2.11% 1.76% 1.52% 1.82% 1.56%
Copper RMSE 0.09 0.10 0.10 0.07 0.09 0.14
MAE 0.07 0.08 0.08 0.05 0.07 0.12
MAPE 1.47% 1.41% 2.02% 1.56% 1.89% 1.44%
Platinum RMSE 19.30 21.95 20.70 21.11 20.17 18.85
MAE 15.22 17.78 16.41 16.47 16.18 14.92
MAPE 1.69% 1.73% 1.73% 1.57% 1.71% 3.33%
Palladium RMSE 89.06 76.33 108.45 70.62 81.21 77.57
MAE 69.72 57.91 79.11 53.54 61.79 59.81
MAPE 4.09% 3.00% 5.20% 2.58% 3.01% 4.08%
Aluminium RMSE 58.54 52.53 56.92 67.75 51.03 49.10
MAE 44.89 41.37 45.92 56.92 40.63 37.55
MAPE 1.59% 1.76% 1.94% 1.70% 1.72% 1.69%

Table 2
RMSE, MAE and MAPE Scores for different models across metals for sub sample data (2003–13) for 60-days’ input.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Stacked LTSM Convolutional-LSTM SVR Bidirectional LSTM XG Boost GRU

Gold RMSE 14.00 53.81 56.92 70.62 17.91 29.00

MAE 10.47 47.92 45.92 53.54 14.27 26.00
MAPE 3.00% 5.50% 0.97% 0.73% 0.84% 1.84%
Silver RMSE 0.59 1.02 0.62 0.64 0.80 0.57
MAE 0.46 0.87 0.48 0.50 0.61 0.45
MAPE 1.74% 1.43% 1.52% 2.58% 1.89% 1.62%
Copper RMSE 0.11 0.04 0.05 0.05 0.04 0.05
MAE 0.10 0.03 0.04 0.04 0.03 0.04
MAPE 1.08% 0.84% 1.15% 3.08% 1.02% 2.12%
Platinum RMSE 19.28 23.73 26.06 19.75 20.79 36.75
MAE 15.40 19.06 21.06 16.61 16.90 31.47
MAPE 1.89% 5.35% 1.30% 0.94% 1.05% 5.28%
Palladium RMSE 32.68 13.87 28.21 15.50 13.35 22.46
MAE 30.70 10.81 21.81 13.03 10.41 19.94
MAPE 2.26% 1.49% 3.34% 1.36% 1.58% 1.96%

Table 3
RMSE, MAE, and MAPE Scores for different models across metals for sub sample data (2013–23) for 60-days’ input.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Stacked LTSM Convolutional-LSTM SVR Bidirectional LSTM XG Boost GRU

Gold RMSE 26.55 29.82 31.61 47.54 24.36 25.92

MAE 22.94 26.40 26.54 44.09 18.19 22.56
MAPE 0.88% 1.12% 1.52% 0.73% 1.05% 0.77%
Silver RMSE 0.45 0.45 0.68 0.46 0.44 0.54
MAE 0.35 0.35 0.56 0.35 0.34 0.42
MAPE 1.92% 1.79% 2.66% 3.18% 1.64% 1.67%
Copper RMSE 0.14 0.15 0.14 0.06 0.07 0.06
MAE 0.12 0.13 0.11 0.04 0.06 0.03
MAPE 1.30% 1.40% 3.10% 1.86% 1.64% 1.63%
Platinum RMSE 23.70 27.20 21.78 19.06 20.74 19.27
MAE 18.66 21.58 17.32 15.49 16.67 15.05
MAPE 1.66% 1.73% 1.80% 1.59% 1.74% 2.38%
Palladium RMSE 73.66 136.08 89.90 90.33 69.54 59.05
MAE 59.33 120.52 70.36 75.00 56.27 47.98
MAPE 4.47% 2.80% 3.69% 2.80% 2.95% 2.52%

Through this, the model attempts to learn the hierarchical representa­ are more prone to overfitting. To avoid over fitting and to improve
tion of data where the lower layers of the model capture the simple generalization, dropout layers can be added in between the LSTM layers.
patterns and the higher layers of the model capture the complex tem­ Studies show that deep-LSTM models with one LSTM layer built on top
poral patterns by building on the simple patterns. of another can significantly augment the learning capability of neural
Stacked-LSTMs can learn more complex representations of the data networks (Cui et al., 2018). Such models have emerged as a prevalent
as each layers captures different aspect of the input data and therefore approach for accurately predicting stock market prices (Cao et al., 2019)

4
A. Varshini et al. Resources Policy 92 (2024) 105040

Table 4
RMSE, MAE and MAPE Scores for different models across metals for full sample data for 30-days’ input.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Stacked LTSM Convolutional-LSTM SVR Bidirectional LSTM XG Boost GRU

Gold RMSE 17.67 33.16 18.84 18.17 18.47 22.24

MAE 13.64 29.12 14.25 14.01 14.25 18.42
MAPE 0.76% 1.62% 0.79% 0.78% 0.79% 1.03%
Silver RMSE 0.59 0.69 0.43 0.42 0.51 0.47
MAE 0.47 0.59 0.32 0.33 0.38 0.37
MAPE 2.14% 2.78% 1.51% 1.52% 1.76% 1.73%
Copper RMSE 0.07 0.07 0.07 0.06 0.10 0.07
MAE 0.06 0.05 0.06 0.05 0.07 0.05
MAPE 1.52% 1.45% 1.54% 1.33% 1.88% 1.46%
Platinum RMSE 29.32 19.58 19.58 20.76 20.11 26.03
MAE 23.85 15.90 15.68 16.39 16.34 20.88
MAPE 2.51% 1.67% 1.65% 1.73% 1.73% 2.17%
Palladium RMSE 69.52 105.53 92.42 91.29 80.00 68.22
MAE 53.13 81.07 66.97 74.17 59.77 51.97
MAPE 2.58% 4.01% 3.19% 3.60% 2.92% 2.55%
Aluminium RMSE 58.40 54.02 52.10 50.41 49.82 58.31
MAE 46.82 43.17 39.85 40.00 39.40 44.00
MAPE 1.98% 1.84% 1.68% 1.70% 1.66% 1.84%

Table 5
RMSE, MAE and MAPE Scores for different models across metals for sub sample data (2003–13) for 30-days’ input.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Stacked LTSM Convolutional-LSTM SVR Bidirectional LSTM XG Boost GRU

Gold RMSE 75.49 175.67 15.75 16.96 15.99 13.49

MAE 74.28 174.82 12.32 13.70 12.38 10.44
MAPE 4.38% 10.30% 0.72% 0.81% 0.73% 0.62%
Silver RMSE 0.57 0.58 0.56 0.53 0.66 0.75
MAE 0.32 0.48 0.43 0.41 0.52 0.61
MAPE 1.43% 1.49% 1.35% 1.31% 1.65% 1.94%
Copper RMSE 0.14 0.06 0.03 0.04 0.05 0.07
MAE 0.13 0.05 0.03 0.03 0.04 0.06
MAPE 3.73% 1.59% 0.84% 0.86% 1.13% 1.85%
Platinum RMSE 22.53 21.59 20.73 23.06 14.71 24.07
MAE 18.86 17.80 16.79 19.52 11.06 18.43
MAPE 1.16% 1.10% 1.04% 1.21% 1.67% 1.14%
Palladium RMSE 12.28 11.03 12.02 12.92 23.52 19.53
MAE 10.23 9.04 9.45 10.89 18.60 16.67
MAPE 1.52% 1.37% 1.43% 1.63% 1.15% 2.52%

Table 6
RMSE, MAE and MAPE Scores for different models across metals for sub sample data (2013–23) for 30-days’ input.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Stacked LTSM Convolutional-LSTM SVR Bidirectional LSTM XG Boost GRU

Gold RMSE 24.44 19.12 18.54 17.72 21.79 21.20

MAE 20.97 15.39 14.91 13.69 16.71 16.90
MAPE 1.21% 0.88% 0.86% 0.78% 0.96% 0.96%
Silver RMSE 0.47 0.48 0.46 0.44 0.51 0.54
MAE 0.37 0.37 0.36 0.34 0.40 0.41
MAPE 1.78% 1.81% 1.73% 1.63% 1.93% 1.97%
Copper RMSE 0.06 0.06 0.08 0.06 0.08 0.06
MAE 0.04 0.04 0.06 0.04 0.06 0.05
MAPE 1.33% 1.34% 1.77% 1.29% 1.88% 1.42%
Platinum RMSE 32.31 19.35 20.23 19.54 19.77 35.13
MAE 26.71 15.44 15.82 16.02 15.36 31.10
MAPE 2.77% 1.60% 1.65% 1.67% 1.61% 3.25%
Palladium RMSE 63.03 59.34 66.77 58.76 63.60 64.65
MAE 52.02 47.72 53.48 47.35 50.67 52.72
MAPE 2.66% 2.45% 2.76% 2.43% 2.64% 2.72%

and have found applications in various domains, including index, forex, a lower parameter count compared to LSTM, leading to decreased
and commodity price prediction (Sezer et al., 2020). memory consumption and improved computational speed. Empirical
studies have demonstrated that GRUs exhibit superior performance
3.2.3. GRU (Gated Recurrent Unit) compared to conventional RNNs (Chung et al., 2014).
The GRU, proposed by Chung et al. (2014), is a variant of RNN. Like Within the depicted architecture in Fig. 4, the GRU employs a soli­
LSTM, GRU does not include an output gate. Nevertheless, GRU exhibits tary update gate for determining the pertinent information to retain,

5
A. Varshini et al. Resources Policy 92 (2024) 105040

Fig. 1. Normalized prices for commodity futures [Note: Aluminium is not included as the data for Aluminium is only available from 2013 onwards.

Fig. 2. The architecture of LSTM.

along with a reset gate to determine the information to discard (See price-to-sales ratio, return on tangible assets, and price-earnings ratio
equations (7)–(9)). Despite its simpler architecture compared to LSTM, (Yang, 2021).
the GRU model has demonstrated competitive performance and gener­
ally yields superior outcomes. This advantage arises from the presence of 3.2.4. Bi-LSTM (bidirectional Long-Short Term Memory)
gates in GRU, particularly the update and reset gates, which facilitate Bi-LSTM represents a variant of bidirectional RNN. In contrast, the
the selective retention or omission of information. This selective pro­ LSTM cells, as illustrated in Fig. 5, can solely process the data in either
cessing of information proves particularly advantageous when the forward or backward direction.
addressing tasks characterized by longer dependencies. Given its ability to process data simultaneously in both forward and
backward directions, the Bi-LSTM model proves advantageous in
Ut = σ(Wu ∗ (ht− 1 , xt )) (7)
capturing dependencies between past and future prices, thereby facili­
tating accurate predictions of future price movements. This model re­
Rt = σ (Wr ∗ (ht− 1 , xt )) (8)
ceives a sequence of past prices as input and predicts the subsequent
Ht = (1 − Ut ) ∗ Ht− 1 + Ut ∗ tanh(W[Rt ∗ Ht− 1 , Ut ]) (9) price in the sequence. The input sequence is processed through separate
LSTM layers for the forward and backward directions. The final pre­
GRU also addresses the issue of vanishing gradients encountered by diction is obtained by concatenating the outputs of each LSTM layer and
vanilla recurrent neural networks. In such networks, the gradient di­ passing them through a dense layer equipped to deal with nonlinearity.
minishes over time during the backpropagation process, rendering the The Bi-LSTM model has demonstrated outstanding performance in
neural networks untrainable as it becomes too small to significantly various domains, including phoneme classification (Graves and
impact the learning process. GRUs find applications in multivariate time Schmidhuber, 2005), speech recognition (Graves et al., 2013), and
series analysis, as well as in predicting financial indicator values such as language modelling (Sundermeyer et al., 2012). Furthermore, Bi-LSTMs

6
A. Varshini et al. Resources Policy 92 (2024) 105040

Fig. 3. Architecture of Stacked-LSTM.

find application in financial time series modelling for predicting stock

and index prices (Siami-Namini et al., 2019).

3.2.5. ConvLSTM (convolutional long short-term memory)

ConvLSTM represents a variant of RNN that was first proposed by Shi
et al. (2015), designed specifically to handle spatiotemporal data. It
achieves this by combining convolutional layers with long-short term
layers. By incorporating this combination, the model can effectively
capture the temporal dependencies as well as the spatial relationships
inherent in the data. Consequently, ConvLSTM can identify the con­
nections between different variables across various spatial areas, thus
enhancing its ability to analyze complex spatiotemporal patterns.
The ConvLSTM model comprises two main components: Convolu­
tional layers and LSTM layers (refer to Fig. 6). The convolutional layers
function similarly to those in standard convolutional neural networks,
extracting features from the input data using learnable filters. Never­
theless, traditional convolutional neural networks tend to focus on
Fig. 4. Architecture of GRU
extracting temporal features, whereas ConvLSTM models integrate both
convolutional and LSTM layers to extract spatial as well as temporal
features. As a result, ConvLSTM models are widely employed in multi­
variate time series forecasting tasks to investigate the spatial-temporal
correlations inherent in the data (Xiao et al., 2021).

Fig. 5. The architecture of Bidirectional-LSTM.

7
A. Varshini et al. Resources Policy 92 (2024) 105040

Fig. 6. The architecture of ConvLSTM.

3.2.6. SVM (Support Vector Regressor model) straight line. On the other hand, the Polynomial kernel is employed
The SVR is a variant of the SVM algorithm introduced by Drucker when there is a non-linear relationship in the data and the input cannot
et al. (1996) and is commonly employed for regression analysis. Unlike be separated by a single line. The RBF Kernel is the most commonly
standard regression techniques that attempt to fit a line to the data utilized kernel function as it allows for capturing intricate non-linear
points, SVR aims to find a non-linear function for prediction. To relationships in the data. Conversely, the Sigmoid kernel is less
accomplish this, SVR utilizes a mapping of variables to a frequently used in SVR and finds its primary application in modelling
higher-dimensional space, enabling their separation by means of a hy­ neural networks.
perplane. The primary goal of SVR is to determine a hyperplane that
minimizes the errors, within a predetermined threshold (see Fig. 7). 3.2.7. XGBoost
SVR has broad applications in time series forecasting across diverse XGBoost is an ensemble learning method (see Fig. 8). Boosting in­
industries, encompassing the stock market, banking sector, and the volves combining multiple weak learners to create a robust learner by
prediction of volatility (Chen et al., 2023). It is also employed as a training a series of weak models on the errors made by previous models.
trend-based segmentation method. In SVR, data is provided as XGBoost assigns higher weightage to the errors made by the previous
input-output pairs, and predictions can be made for future time steps. model, allowing for subsequent correction and leading to improved
This offers a significant advantage over traditional regression techniques performance and higher accuracy. XGBoost demonstrates a particular
since SVR enables the capture of complex non-linear relationships be­ suitability for time series forecasting tasks as it possesses the capability
tween past and future time steps. to effectively handle temporal dependencies among variables while
In the context of SVR modelling, the choice of a suitable kernel identifying non-linear relationships that exist between them. Compared
function is an important step. There are four kinds of kernel functions to standard ML models, XGBoost offers easier handling of missing data.
available: Linear, Polynomial, Radial Basis Function (RBF), and Sig­ The model is an ensemble of decision trees built in a depth-wise manner,
moid. The Linear kernel, being the simplest, is suitable for cases where a wherein these models are sequentially trained to address errors made by
linear relationship exists or when the input data can be separated by a the previous weak learners. The algorithm incorporates various opti­
mization and regularization techniques, which contribute to its power,
flexibility, and popularity in the field of ML. XGBoost has found appli­
cations in the financial time series domain for forecasting stock market
volatility, predicting commodity prices (Jabeur et al., 2021; Baser et al.,
2023), and imputing time series data.

4. Results and discussion

4.1. Evaluation measures

To assess the performance of these machine learning (ML) and deep

learning (DL) models, a comparative analysis is conducted among
Stacked-LSTM, ConvLSTM, Support Vector Regressor (SVR), Bi-LSTM,
Extreme Gradient Boosting (XGBoost), and Gated Recurrent Unit
(GRU). We employ metrics such as Root Mean Square Error (RMSE),
Mean Absolute Error (MAE), and Mean Absolute Percentage Error
(MAPE) to check for the forecasting performance of these models. These
metrics facilitate model comparison to identifying the most suitable
model. RMSE is one of the most adopted metrics to assess models’ per­
formance. It quantifies the squared deviation between the observed and
predicted values and then calculates the square root of the mean of these
Fig. 7. Illustration of Support Vector Regressor function. squared deviations as shown in equation (10) (Astudillo et al., 2020;

8
A. Varshini et al. Resources Policy 92 (2024) 105040

Fig. 8. Architecture of XGBoost.

Baser et al., 2023). MAE, RMSE, and MAPE each possess their own set of advantages and
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ disadvantages. Consequently, it is imperative to employ these metrics
√ n
√∑ (y − ˙ yi ) collectively to achieve a comprehensive assessment of the models’
RMSE = √ (10)
i=1
n performances (Baser et al., 2023).

One of the major reasons for choosing RMSE as the evaluation metric
because it serves as a loss function for optimization in the neural net­ 4.2. Findings and discussion
works that we have trained. Therefore, it allows us to directly evaluate
the predictive quality across different models. Furthermore, RMSE ap­ The study discovered that RMSE, MAE, and MAPE yield similar re­
plies a greater penalty to large errors, thereby emphasizing the impor­ sults regarding the performance of the models in most of the cases, as
tance of accurately capturing deviations from the actual values. For time shown in Table 1.
series forecasting performance, a lower RMSE score is better. The RMSE, and MAE values for the following metals (Gold, Platinum,
MAE is an alternative metric that offers greater robustness to outliers Palladium, and Aluminium) are high. The lower MAPE values for gold
compared to RMSE, as it does not involve squaring the errors. MAE (0.74%), and aluminium (1.59%) denote superiority of stacked LSTM
computes the mean absolute discrepancy between the observed and over other models for forecasting their future prices. Similarly, Bidi­
predicted values, providing a straightforward measure of the overall rectional LSTM yields better forecasting estimates for silver (1.52%),
discrepancy between them as shown in equation (11). platinum (1.57%), and palladium (2.58%). Then, Convolution LSTM
finds to be better forecasting model for forecasting the copper (1.41%)
1 ∑n
future prices. Notably, the GRU finds to be the second-best performing
MAE = | y − yi | (11)
n i=1 models for forecasting the future prices of following metals namely gold
(0.79%), silver (1.56%), copper (1.44%), and aluminium (1.69%). The
MAE is an easily interpretable metric that is less affected by outliers,
SVR model finds to be the least performing for copper (2.02%), palla­
as it does not penalize extreme values as severely as RMSE does (Chai
dium (5.20%), and aluminium (1.94%).
and Draxler, 2014). A lower MAE indicates better accuracy, as it sig­
This implies that the best or least-performing models differ
nifies that the forecasted values are, on average, closer to the actual
depending on the specific metal being analyzed. Given the varying re­
values.
sults exhibited by different ML models for different metals within the
Along with MAE and RMSE, we also consider MAPE. It is a metric
same sample period, we proceed to conduct a sub-sample analysis to
used to evaluate the accuracy of forecasts by measuring the average
assess whether this variation persists within the same metal but with a
percentage difference between the forecasted values and the corre­
different sample period. To accomplish this, we divide the full sample
sponding actual values. MAPE gives a sense of how well a forecasting
data into two equal periods: February 2003 to January 2013 and
model’s predictions align with the actual data in terms of percentage
February 2013 to January 2023. This division is motivated by the
deviations as shown in equation (12).
presence of a structural break in the future prices of metals during early
1 ∑n
y − yi 2013. However, due to limited aluminium future price data, we exclude
MAPE = ∗ 100 (12)
n i=1 y aluminium from the sub-sample analysis.
Table 2 displays the outcomes of the initial sub-sample analysis. Here
MAPE is expressed as a percentage, and a lower MAPE value in­ some estimates differ from the full sample estimations as follows. The
dicates better accuracy. However, it is important to note that MAPE can Bidirectional LSTM model MAPE value finds to be best for forecasting
encounter issues when dealing with small or zero actual values, as the gold future prices. Whereas Convolutional LSTM MAPE value shows the
percentage difference might become undefined or excessively large. worst performance for forecasting gold future prices among all the

9
A. Varshini et al. Resources Policy 92 (2024) 105040

models. Similarly, for forecasting the silver future prices the estimates consideration in forecasting the future metal prices. It also confirms that
show vice-versa to gold estimates. Then for forecasting the copper future each model’s performance under consideration on forecasting the future
prices Bidirectional LSTM model finds to be the worst among all the prices of the metals greatly varies with choice of metals, period, and
models. These findings highlight the variation in model performance inputs. While metals like gold and silver are considered safe-haven as­
across different metals and different data samples. No single model sets offering portfolio diversification, their speculative nature cannot be
consistently outperforms the others for any metal. Next, we aim to ignored (Janani Sri et al., 2022). Additionally, the price variations
investigate whether similar observations persist in the second sub- among different metals may not be directly linked, but they often move
sample. in the same direction, which can impact price prediction and model
Table 3 exhibits the second sub-sample analysis results. In the second performance. Explaining the variations in model performance empiri­
sub-sample, we observe relatively more variation in relation to the cally is a challenge, as they can be influenced by multiple factors.
performance of the model with the full sample as follows. Based on the The key reasons for such variability or uncertainty in the model
MAPE values Bidirectional LSTM (0.73%), XG Boost (1.64%), Stacked performances is due to the following. First, timeseries data distribution
LSTM (1.30%), and GRU (2.52%) find to be the best-performing models changes over a period. Second, variability in data distribution results in
for forecasting future prices of the metal’s gold, silver, copper, and concept drift and structural breaks. Due to which there are deviations in
palladium respectively. Overall, our observations indicate three characteristics of the underlying relationship between features and
important findings: (i) the best-performing model varies across metals, target variable over time. Third, dissimilar input features might be
(ii) the choice of data period significantly impacts the models’ perfor­ relevant during different time periods. Fourth, there could be demand
mance, and (iii) inputs to the models.’ These findings imply that there is and supply variations in different metals. Fifth, different models have
no consensus on the best model for predicting metals’ future prices. their own strengths and weaknesses that affect its overall performance.
Assigning higher importance to recent data when utilizing these models Sixth, various external factors such as geopolitical, technological, legal,
for price prediction may lead to improved performance. However, none and other spheres affect the metal prices. Seventh, another important
of the particular model able to maintain constant performance over time factor is “tail risk,” which most ML models do not appropriately account
for predicting metals’ future prices except in the case of forecasting the for. As a result, they may underestimate the potential for extreme events,
future prices for platinum (60 days inputs). Irrespective of the sample leading to inaccurate predictions during periods of heightened volatility
periods, Bidirectional LSTM, and GRU models are the best and worst or market stress. The presence of non-linearity, non-stationarity, and
performance models for forecasting the future prices for platinum with volatility in price series introduces complexity, posing challenges in
60 days inputs. However, the model performance varies significantly consistently and accurately predicting metal prices, even when utilizing
with 30 days input even in case of platinum. This indicates that market ML and DL models (Luo et al., 2022). Lastly, varying model performance
efficiency is not a binary concept but exists along a spectrum. Thus, across metals can also be linked to investor sentiment and biases specific
different assets may exhibit varying degrees of efficiency with time. This to different metals. The study findings have high relevance for portfolio
inherent challenge in making accurate predictions suggests that market theory, hedging and risk Management. The variation in the model per­
participants would find it difficult to consistently outperform the market formance across metals can influence portfolio construction and asset
in predicting metals’ future prices. Hence, our observations provide allocation decisions. Consequently, it can have a direct influence on the
empirical support for the efficient market hypothesis (Fama, 1970). That effectiveness of hedging strategies and risk management techniques.
suggests that it is hard to beat the market constantly with a single Thus, to address these issues continuous innovation of models is essen­
strategy. The forecasts for all the considered series using different tial to enhance the accuracy and robustness of forecasting.
techniques are visualized in Appendix II. The estimates with 30 days’
inputs are shown in Tables 4–6. The obtained estimates show that the 5. Conclusion
best and worst-performing models for forecasting metal future prices
vary significantly from 60 days as inputs. It confirms that ML and DL Commodity markets play a crucial role in the global economy as they
models’ performance can vary based on sample periods and inputs due provide essential raw materials for various industries, with a significant
to a combination of factors related to data, model complexity, and the portion of commodities traded on the futures market (Chang and Fang,
underlying characteristics of the problem being solved. 2022). Among the commodities, metals are particularly important and
The consistency of the best-performing model across different pe­ often considered safe-haven assets during economic crises (Peng, 2021).
riods and choice of metals is not evident in our analysis as shown in This study focuses on forecasting metal futures (e.g., gold, silver, copper,
Fig. 9. It shows that in most of the cases MAPE values are less than 5% platinum, palladium, and aluminium) in commodity markets using
that overall confirm the superiority of ML and DL models under different machine learning (ML) and deep learning (DL) techniques. The

Fig. 9. MAPE values of the different metals across the various metals with different inputs
Note: Metal with no subscript represents full sample; subscript_1 represent subsample (2003–2013); and subscript_2 represent subsample (2013–2023).

10
A. Varshini et al. Resources Policy 92 (2024) 105040

models utilized in this study comprises of the Stacked-LSTM, Convolu­ ways to enhance our understanding of this dynamic field. Combining
tional LSTM, Support Vector Regressor (SVR), Bidirectional LSTM, technical and fundamental analysis can provide a more holistic view of
Extreme Gradient Boosting (XGBoost), and Gated Recurrent Unit (GRU). commodity market behaviour, incorporating factors like supply and
The analysis is conducted in two parts. First, the models are trained on demand, geopolitical influences, technological, legal, and macroeco­
20 years of data, from February 2003 to January 2023. Second, the nomic indicators. Therefore, building or validating any theory based on
dataset is further bifurcated into two subsamples each of ten years for the ML and DL models estimations alone is not advisable. The tradi­
robust analysis. The models’ performance is assessed by employing tional, ML, and DL models differ in purpose, focus, and methodologies.
evaluation metrics such as Root Mean Squared Error (RMSE), Mean The future studies may consider hybrid models by the integration of
Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). different traditional, ML, and DL for forecasting future prices of the
Most of the research in this domain concentrates on forecasting the metals. The future studies may also investigate the application of deep
prices of individual metals (Vidya and Hari, 2020; Yamak et al., 2019), reinforcement learning (DRL) for trading strategies in commodity mar­
and only a few have performed sub-sample analysis due to limited data kets holds promise for adaptive and responsive trading approaches.
availability. This study stands out by considering a unique combination
of metal assets and conducting sub-sample analysis, including data CRediT authorship contribution statement
normalization and identification of structural breakpoints to determine
the sub-sample periods. Additionally, the present study uses different Anu Varshini: Conceptualization, Data curation, Formal analysis,
inputs periods of 30 and 60 days respectively for robust analysis. The Investigation, Methodology, Software, Validation, Visualization,
RMSE, MAE, and MAPE scores indicate that the performance of models Writing – original draft. Parthajit Kayal: Conceptualization, Formal
varied across different metals, inputs, and periods indicating the absence analysis, Project administration, Supervision, Validation, Writing – re­
of a consistently superior model. The results support the efficient market view & editing. Moinak Maiti: Conceptualization, Formal analysis,
hypothesis and indicate that consistently outperforming the market in Supervision, Writing – review & editing.
predicting metal prices may be challenging. Given the heterogeneity in
predictive performance across different metals, and time periods, mar­ Declaration of competing interest
ket participants, including traders and investors, should consider
diversifying their investment strategies. The study findings show that There is no conflict of interest.
conducting comprehensive sub-sample and rolling window analyses
(Kayal and Balasubramanian, 2021; Som and Kayal, 2022) is essential Acknowledgement
before establishing the reliability of various ML and DL models for ac­
curate price predictions. Authors are thankful to the University of Witwatersrand, Johannes­
Future studies in commodity price forecasting should explore several burg, South Africa for supporting Open Access Publications.

Appendix-I
Table A
Parameters settings of Stacked LTSM, Convolutional-LSTM, Bidirectional LSTM, and GRU

Parameters/Hyperparameters Stacked LTSM Convolutional-LSTM Bidirectional LSTM GRU

filters 128
kernel_size (3, 3)
strides (1, 1)
padding same’
activation tanh’
units 128 128 128 128
activation tanh tanh’ tanh tanh
recurrent_activation sigmoid sigmoid’ sigmoid sigmoid
use_bias TRUE TRUE TRUE TRUE
kernel_initializer glorot_uniform glorot_uniform glorot_uniform glorot_uniform
recurrent_initializer orthogonal orthogonal orthogonal orthogonal
bias_initializer zeros zeros zeros zeros
kernel_regularizer None None None None
recurrent_regularizer None None None None
bias_regularizer None None None None
kernel_constraint None None None None
recurrent_constraint None None None None
bias_constraint None None None None
dropout 0 0 0
recurrent_dropout 0 0 0

Table B
Parameters settings of SVR and XGBoost Models

XG Boost SVR

Parameter/Hyperparameter Value Parameter/Hyperparameter Value

objective reg: squarederror kernel rbf
booster None C (Penalty Parameter) 10
n_estimators 1000 epsilon 0.01
learning_rate 0.01 gamma 0.8
(continued on next page)

11
A. Varshini et al. Resources Policy 92 (2024) 105040

Table B (continued )
XG Boost SVR

max_depth 3
min_child_weight 1
subsample 1
colsample_bytree 1

Appendix-II

Fig. A. Predictions of prices of Gold, Silver, Copper, Platinum, Palladium and Aluminium through Stacked-LSTM Model.

12
A. Varshini et al. Resources Policy 92 (2024) 105040

Fig. B. Predictions of prices of Gold, Silver, Copper, Platinum, Palladium and Aluminium through ConvLSTM Model.

Fig. C. Predictions of prices of Gold, Silver, Copper, Platinum, Palladium and Aluminium through SVR Model.

13
A. Varshini et al. Resources Policy 92 (2024) 105040

Fig. D. Predictions of prices of Gold, Silver, Copper, Platinum, Palladium and Aluminium through Bi-LSTM Model.

Fig. E. Predictions of prices of Gold, Silver, Copper, Platinum, Palladium and Aluminium through XGBoost Model.

14
A. Varshini et al. Resources Policy 92 (2024) 105040

Fig. F. Predictions of prices of Gold, Silver, Copper, Platinum, Palladium and Aluminium through GRU Model.

References Chai, T., Draxler, R.R., 2014. Root mean square error (RMSE) or mean absolute error
(MAE)?–Arguments against avoiding RMSE in the literature. Geoscien. model
develop 7 (3), 1247–1250.
Ampomah, E.K., Qin, Z., Nyame, G., 2020. Evaluation of tree-based ensemble machine
Chang, C.L., Fang, M., 2022. The connectedness between natural resource commodities
learning models in predicting stock price direction of movement. Information 11 (6),
and stock market indices: evidence from the Chinese economy. Res. Policy 78,
332.
102841.
Arowolo, W.B., 2013. Predicting stock prices returns using GARCH model. Int. J. Eng.
Chen, W.J., Yao, J.J., Shao, Y.H., 2023. Volatility forecasting using deep neural network
Sci. 2 (5), 32–37.
with time-series feature embedding. Econ. Resch.-Ekonomska Istraživanja 36 (1),
Astudillo, G., Carrasco, R., Fernández-Campusano, C., et al., 2020. Copper price
1377–1401.
prediction using support vector regression technique. Appl. Sci. 10 (19), 6648.
Chen, Z., Goh, H.S., Sin, K.L., et al., 2021. Automated agriculture commodity price
Baser, P., Saini, J.R., Baser, N., 2023. Gold commodity price prediction using tree-based
prediction system with machine learning techniques. arXiv preprint arXiv:
prediction models. Int. J. Intel. Syst. Appl. Eng. 11 (1s), 90–96.
2106.12747.
Bhardwaj, S.P., Paul, R.K., Singh, D.R., et al., 2014. An empirical investigation of ARIMA
Chung, J., Gulcehre, C., Cho, K., et al., 2014. Empirical Evaluation of Gated Recurrent
and GARCH models in agricultural price forecasting. Econ. Aff. 59 (3), 415–428.
Neural Networks on Sequence Modeling. arXiv Preprint arXiv:1412.3555.
Bianchi, D., Büchner, M., Tamoni, A., 2021. Bond risk premiums with machine learning.
Cui, Z., Ke, R., Pu, Z., et al., 2018. Deep bidirectional and unidirectional LSTM recurrent
Rev. Financ. Stud. 34 (2), 1046–1089.
neural network for network-wide traffic speed prediction. arXiv preprint arXiv:
Cao, J., Li, Z., Li, J., 2019. Financial time series forecasting model based on CEEMDAN
1801.02143.
and LSTM. Physica A 519, 127–139.
Desai, V.S., Bharati, R., 1998. A comparison of linear regression and neural network
Cao, L.J., Tay, F.E.H., 2003. Support vector machine with adaptive parameters in
methods for predicting excess returns on large stocks. Anl. Oper. Resch. 78 (0),
financial time series forecasting. IEEE Trans. Neural Network. 14 (6), 1506–1518.
127–163.
Drucker, H., Burges, C.J., Kaufman, L., et al., 1996. Support vector regression machines.
Adv. Neural Inf. Process. Syst. 9.

15
A. Varshini et al. Resources Policy 92 (2024) 105040

Essien, A., Giannetti, C., 2019. A deep learning framework for univariate time series Onour, I., Sergi, B.S., 2011. Modeling and forecasting volatility in the global Food
prediction using convolutional LSTM stacked autoencoders. In: 2019 IEEE Int. Symp. commodity prices. Agri. Econ. -Czech 57 (3), 132–139.
Innov. Intelligent SysTem. IEEE, pp. 1–6. Appl. (INISTA). Parmezan, A.R.S., Souza, V.M., Batista, G.E., 2019. Evaluation of statistical and machine
Fama, E.F., 1970. Efficient capital markets: a review of theory and empirical work. learning models for time series prediction: identifying the state-of-the-art and the
J. Finance 25 (2), 383–417. best conditions for the use of each model. Inf. Sci. 484, 302–337.
Fama, E.F., 1984. Forward and spot exchange rates. J. Monetary Econ. 14 (3), 319–338. Peng, Z., 2021. Precious Metal Price Prediction Using Deep Neural Networks.
Gao, L., Guo, Z., Zhang, H., et al., 2017. Video captioning with attention-based LSTM and Ranjan, S., Kayal, P., Saraf, M., 2023. Bitcoin price prediction: a machine learning sample
semantic consistency. IEEE Trans. Multimed. 19 (9), 2045–2055. dimension approach. Comput. Econ. 61, 1617–1636.
Gong, X., Xu, J., 2022. Geopolitical risk and dynamic connectedness between commodity Raza, S.A., Masood, A., Benkraiem, R., Urom, C., 2023. Forecasting the volatility of
markets. Energy Econ. 110, 106028. precious metals prices with global economic policy uncertainty in pre and during the
Graves, A., Mohamed, A.R., Hinton, G., 2013. Speech recognition with deep recurrent COVID-19 period: novel evidence from the GARCH-MIDAS approach. Energy Econ.
neural networks. In: 2013 IEEE International Conference on Acoustics, Speech and 120, 106591.
Signal Processing. IEEE, pp. 6645–6649. Reddy, V.K.S., 2018. Stock market prediction using machine learning. Int. Res. J. Eng.
Graves, A., Schmidhuber, J., 2005. Framewise phoneme classification with bidirectional Tech. 5 (10), 1033–1035.
LSTM and other neural network architectures. Neural Network. 18 (5–6), 602–610. Schnepf, R.D., Resources, Science, Division, Industry, 2005. Price determination in
Han, S., Kang, J., Mao, H., Hu, Y., et al., 2017. Ese: efficient speech recognition engine agricultural commodity markets: a primer. Cong. Res. Serv. Library of Congress.
with sparse lstm on fpga. In: Proceedings of the 2017 ACM/SIGDA International (December).
Symposium on Field-Programmable Gate Arrays, pp. 75–84. Sezer, O.B., Gudelek, M.U., Ozbayoglu, A.M., 2020. Financial time series forecasting with
Herrera, G.P., Constantino, M., Tabak, B.M., et al., 2019. Long-term forecast of energy deep learning: a systematic literature review: 2005–2019. App. Soft Comput. 90,
commodities price using machine learning. Energy 179, 214–221. 106181.
Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neu. Comp. 9 (8), Shahwan, T., Odening, M., 2007. Forecasting agricultural commodity prices using hybrid
1735–1780. neural networks. Comp. Intel. Econ. Fin. 2, 63–74.
Horák, J., Jannová, M., 2023. Predicting the Oil price movement in commodity markets Sherstinsky, A., 2020. Fundamentals of recurrent neural network (RNN) and long short-
in global economic Meltdowns. Forecasting 5 (2), 374–389. term memory (LSTM) network. Phys. Nonlinear Phenom. 404, 132306.
Hussein, S.F.M., Shah, M.B.N., Abd Jalal, M.R., et al., 2011. Gold price prediction using Shi, X., Chen, Z., Wang, H., et al., 2015. Convolutional LSTM network: a machine
radial basis function neural network. In: 2011 Fourth International Conference on learning approach for precipitation nowcasting. Adv. Neural Inf. Process. Syst. 28.
Modeling, Simulation and Applied Optimization. IEEE, pp. 1–11. Siami-Namini, S., Tavakoli, N., Namin, A.S., 2019. A comparative analysis of forecasting
Jabeur, S.B., Mefteh-Wali, S., Viviani, J.L., 2021. Forecasting gold price with the financial time series using arima, lstm, and bilstm. arXiv preprint arXiv:1911, 09512.
XGBoost algorithm and SHAP interaction values. Anl. Oper. Resch 1–21. Som, A., Kayal, P., 2022. A multicountry comparison of cryptocurrency vs gold: portfolio
Jacks, D.S., O’rourke, K.H., Williamson, J.G., 2011. Commodity price volatility and optimization through generalized simulated annealing. BLOCK: Res. Appl. 3 (3),
world market integration since 1700. Rev. Econ. Stat. 93 (3), 800–813. 100075.
Janani Sri, S., Kayal, P., Balasubramanian, G., 2022. Can equity be safe-haven for Sundermeyer, M., Schlüter, R., Ney, H., 2012. LSTM neural networks for language
investment? J. Emerg. Mark. Finance 21 (1), 32–63. modeling. In: Thirteenth Annual Conference of the International Speech
Kayal, P., Balasubramanian, G., 2021. Excess volatility in bitcoin: extreme value Communication Association.
volatility estimation. IIM Kozhikode Soc. Manag. Rev. 10 (2), 222–231. Torres, D.G., Qiu, H., 2018. Applying Recurrent Neural Networks for Multivariate Time
Kayal, P., Maheswaran, S., 2021. A study of excess volatility of gold and silver. IIMB Series Forecasting of Volatile Financial Data. KTH Royal Institute of Technology,
Manag. Rev. 33 (2), 133–145. Stockholm, Sweden.
Kilinc, H.C., Yurtsever, A., 2022. Short-term streamflow forecasting using hybrid deep Valencia, F., Gómez-Espinosa, A., Valdés-Aguirre, B., 2019. Price movement prediction
learning model based on grey wolf algorithm for hydrological time series. of cryptocurrencies using sentiment analysis and machine learning. Entropy 21 (6),
Sustainability 14 (6), 3352. 589.
Krollner, B., Vanstone, B.J., Finnie, G.R., 2010. Financial time series forecasting with Verheyen, F., 2010. Monetary Policy, Commodity Prices and Inflation-Empirical
machine learning techniques: a survey. In: ESANN (April. Evidence from the US. Commodity Prices and Inflation-Empirical Evidence from the
Kumar, K.A., Spulbar, C., Pinto, P., et al., 2022. Using econometric models to manage the US (November, 2010).
price risk of Cocoa Beans: a case from India. Risks 10 (6), 115. Vidya, G.S., Hari, V.S., 2020. Gold price prediction and modelling using deep learning
Li, L., Wu, Y., Ou, Y., et al., 2017. Research on machine learning algorithms and feature techniques. In: 2020 IEEE Recent Advances in Intelligent Computational Systems
extraction for time series. In: 2017 IEEE 28th Annual International Symposium on (RAICS). IEEE, pp. 28–31.
Personal, Indoor, and Mobile Radio Communications (PIMRC). IEEE, pp. 1–5. Vukovic, D., Vyklyuk, Y., Matsiuk, N., et al., 2020. Neural network forecasting in
Li, Y., Bao, T., Gong, J., et al., 2020. The prediction of dam displacement time series prediction Sharpe ratio: evidence from EU debt market. Physica A 542, 123331.
using STL, extra-trees, and stacked LSTM neural network. IEEE Access 8, Xiao, Y., Yin, H., Zhang, Y., et al., 2021. A dual-stage attention-based Conv-LSTM
94440–94452. network for spatio-temporal correlation and multivariate time series prediction. Int.
Lim, B., Zohren, S., 2021. Time-series forecasting with deep learning: a survey. Philosop. J. Intell. Syst. 36 (5), 2036–2057.
Transac. Royal Societ. A 379 (2194), 20200209. Yamak, P.T., Yujian, L., Gadosey, P.K., 2019. A comparison between arima, lstm, and gru
Livieris, I.E., Pintelas, E., Pintelas, P., 2020. A CNN–LSTM model for gold price time- for time series forecasting. In: Proceedings of the 2019 2nd International Conference
series forecasting. Neural Comput. Appl. 32, 17351–17360. on Algorithms, Computing and Artificial Intelligence, pp. 49–55.
Luo, H., Wang, D., Cheng, J., et al., 2022. Multi-step-ahead copper price forecasting using Yang, S., 2021. A novel study on deep learning framework to predict and analyze the
a two-phase architecture based on an improved LSTM with novel input strategy and financial time series information. Future Generat. Comput. Syst. 125, 812–819.
error correction. Res. Policy 79, 102962. Zhang, H., Nguyen, H., Vu, D.A., et al., 2021a. Forecasting monthly copper price: a
Maiti, M., Vukovic, D., Vyklyuk, Y., et al., 2022. BRICS capital markets Co-movement comparative study of various machine learning-based methods. Res. Pol. 73,
analysis and forecasting. Risks 10 (5), 88. 102189.
Maiti, M., Vyklyuk, Y., Vuković, D., 2020. Cryptocurrencies chaotic co-movement Zhang, L., Bian, W., Qu, W., et al., 2021b. Time series forecast of sales volume based on
forecasting with neural networks. Intern. Tech. Letters 3 (3), e157. XGBoost. In Journal of Physics: Conference Series 1873 (No. 1), 012067. IOP
Makala, D., Li, Z., 2021. Prediction of gold price with ARIMA and SVM. In Journal of Publishing.
Physics: Conference Series. 1767 (1), 012022. Zhou, J., Cao, Y., Wang, X., et al., 2016. Deep recurrent models with fast-forward
Nikou, M., Mansourfar, G., Bagherzadeh, J., 2019. Stock price prediction using DEEP connections for neural machine translation. Trans. Asso. Comp. Ling. 4, 371–383.
learning algorithm and its comparison with machine learning algorithms. Intell. Zhou, J., He, Z., Song, Y.N., et al., 2019. Precious metal price prediction based on deep
Syst. Account. Finance Manag. 26 (4), 164–174. regularization self-attention regression. IEEE Access 8, 2178–2187.
Ojo, S.O., Owolawi, P.A., Mphahlele, M., et al., 2019. Stock market behaviour prediction Zulauf, C.R., Irwin, S.H., Ropp, J.E., et al., 1999. A reappraisal of the forecasting
using stacked LSTM networks. In: 2019 International Multidisciplinary Information performance of corn and soybean new crop futures. J. Futures Mark. 19 (5),
Technology and Engineering Conference (IMITEC). IEEE, pp. 1–5. 603–618.

16