Journal Pre-proof

A hybrid deep learning approach by integrating LSTM-ANN networks

with GARCH model for copper price volatility prediction

Yan Hu, Jian Ni, Liu Wen

PII: S0378-4371(20)30469-6
DOI: https://doi.org/10.1016/j.physa.2020.124907
Reference: PHYSA 124907

To appear in: Physica A

Received date : 1 December 2019

Revised date : 7 June 2020

Please cite this article as: Y. Hu, J. Ni and L. Wen, A hybrid deep learning approach by integrating
LSTM-ANN networks with GARCH model for copper price volatility prediction, Physica A
(2020), doi: https://doi.org/10.1016/j.physa.2020.124907.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the
addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive
version of record. This version will undergo additional copyediting, typesetting and review before it
is published in its final form, but we are providing this version to give early visibility of the article.
Please note that, during the production process, errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.

© 2020 Published by Elsevier B.V.

 Journal Pre-proof

A hybrid deep learning approach by integrating LSTM-ANN networks with

GARCH model for copper price volatility prediction

of
Yan Hu

Tel.: +86 28 87352835

E-mail: [email protected]

pro
School of Finance, Southwestern University of Finance and Economics

Chengdu 611130, China

re-
** Jian Ni

(Corresponding author)

Tel.: +86 28 87352835

E-mail: [email protected]
lP

School of Finance, Southwestern University of Finance and Economics

Chengdu 611130, China

rna

Liu Wen
Tel.: +86 28 87352835

E-mail: [email protected]

School of Finance, Southwestern University of Finance and Economics

Jou

Chengdu 611130, China

 Journal Pre-proof

A hybrid deep learning approach by integrating LSTM-ANN networks with

GARCH model for copper price volatility prediction

Abstract: Forecasting the copper price volatility is an important yet challenging task. Given

of
the nonlinear and time-varying characteristics of numerous factors affecting the copper price,

we propose a novel hybrid method to forecast copper price volatility. Two important techniques

are synthesized in this method. One is the classic GARCH model which encodes useful

pro
statistical information about the time-varying copper price volatility in a compact form via the

GARCH forecasts. The other is the powerful deep neural network which combines the GARCH

forecasts with both domestic and international market factors to search for better nonlinear

features; it also combines the long short-term memory (LSTM) network with traditional
re-
artificial neural network (ANN) to generate better volatility forecasts. Our method synthesizes

the merits of these two techniques and is especially suitable for the task of copper price

volatility prediction. The empirical results show that the GARCH forecasts can serve as

informative features to significantly increase the predictive power of the neural network model,
lP

and the integration of the LSTM and ANN networks is an effective approach to construct useful

deep neural network structures to boost the prediction performance. Further, we conducted a

series of sensitivity analyses of the neural network architecture to optimize the prediction
rna

results. The results suggest that the choice between LSTM and BLSTM networks for the hybrid

model should consider the forecast horizon, while the ANN configurations should be fine-tuned

depending on the choice of the measure of prediction errors.

Keywords: copper price volatility; deep learning; GARCH; LSTM-ANN

Jou

1. Introduction

As a key material in various industrial applications, copper is the most actively traded base

metal. However, copper prices are highly volatile and depend on many external factors. Such

inherent high volatility makes the prediction modelling more challenging. In fact, the

application of traditional prediction models (e.g., ARMA and GARCH) has shown their

underperformance on the volatility forecasts of copper prices. Despite the challenges,

improving the volatility forecasts of copper prices is a worthy endeavor since volatility is an
1
 Journal Pre-proof

important proxy for market risk. In particular, Kristjanpoller and Minutolo (2015) and García

and Kristjanpoller (2019) have adopted volatility for measuring the risk of commodity market.

Ederington and Lee (1993) and Fuertes et al. (2009) have also applied volatility to measure the

risk in the stock market. Other applications of volatility as a measure of market risk can be

of
found in Kristjanpoller et al. (2014), Kristjanpoller and Hernández (2017), Kim and Won (2018),

and references therein. Therefore, the ability to predict the volatility of copper prices with

greater precision is critical for market participants. How to accurately predict volatility is still

pro
an unsolved issue. To address this issue, a hybrid volatility prediction model is developed in

this study by synthesizing the state-of-art deep learning technique with the classic GARCH

(Generalized Autoregressive Conditional Heteroskedasticity) model.

In the finance industry, volatility is usually predicted by various GARCH-type models. As

re-
a useful generalization of the autoregressive conditional heteroskedasticity (ARCH) model, the

GARCH model proposed by Bollerslev (1986) provides a way to model a change in variance

in a time series. Later on, different extensions on the basic ARCH and GARCH models have

been developed, including EGARCH (Nelson, 1991), APGARCH (Ding et al., 1993),
lP

SWARCH (Hamilton and Susmel, 1994), FIAPGARCH (Tse, 1998), HYGARCH (Davidson,

2004), and more. However, due to the existence of complex nonlinear correlation structure

among variables and larger mass of data sets, the prediction results of these GARCH-type
rna

models may not be very satisfactory.

The recent development of the deep learning methodology originates from the artificial

neural network (ANN) models, which are designed to mimic the knowledge-acquisition and

organizational skills of the human brain (Bergerson and Wunsch, 1991; Sharda and Patil, 1992).
Jou

Many comparative studies between ANNs and traditional prediction models (e.g., ARMA and

GARCH) have been conducted with regard to their performance on predictions. High accuracy

of ANNs on volatility prediction for various commodities have been demonstrated, see Hamid

and Iqbal (2004), HParisi et al. (2008), Azadeh et al. (2012) and Yazdani-Chamzini et al. (2012).

Furthermore, hybrid ANN and GARCH-type models are usually found to have advantages

in comparison with ANNs or time series models. Bildirici and Ersin (2013) used a neural

network augmented GARCH to predict oil prices. They concluded that neural networks models

were promising. Kristjanpoller et al. (2014) applied a hybrid ANN-GARCH model to forecast
2
 Journal Pre-proof

volatility in three Latin American indexes from Brazil, Chile, and Mexico. They demonstrated

that neural network models can improve the predictions from GARCH models. Similar studies

about hybrid models with ANNs can be seen as Kristjanpoller and Minutolo (2015), Cui et al.

(2015), Lu et al. (2016), Lahmiri (2017) and Kristjanpoller and Hernández (2017).

of
Extensions of ANNs also improve the accuracy of copper price prediction. A recurrent

neural network (RNN) is one of the extensions of ANNs. In the RNN, connections between

nodes form a directed graph along a temporal sequence. This allows the neural network to

pro
exhibit temporal dynamic behavior. Unlike feedforward neural networks, RNNs can use their

internal state (memory) to process sequences of inputs. This makes them applicable to tasks

such as unsegmented, connected handwriting recognition (Graves et al., 2009) or speech

recognition (Li and Wu, 2015). Various forms of RNNs have performed well for prediction as
re-
related to time series data (Herman and Schrauwen, 2013), such as the long short-term memory

(LSTM) and bidirectional LSTM (BLSTM). The LSTM network, which was firstly introduced

by Hochreiter and Schmidhuber (1997), is widely perceived as a successful RNN model in

dealing with the time series data, see Maknickienė and Maknickas (2012), Nelson et al. (2017)
lP

and Kim et al. (2018), Wu et al (2019). BLSTM is proposed to access both the past and future

information by combining a forward hidden layer and a backward hidden layer (Liu and Guo,

2019). Sardelich and Manandhar (2018) used a model combining the BLSTM and stacked
rna

LSTM neural networks to predict daily stock volatility. The proposed model outperformed the

well-known GARCH (1,1) model in many sectors (e.g., financial, health care, etc.). Eapen et

al. (2019) found that CNN layers combining with BLSTM showed a better prediction

performance than the traditional SVM regressor.

Jou

Therefore, the recent advancement of the AI techniques, characterized by the so-called

deep learning methodology, opens up possibilities for us to develop new prediction models that

can better forecast the volatility of copper prices. To enhance the prediction power, the best

forecasts of the GARCH model are used as inputs for the hybrid models that combine GARCH

and neural networks. Six neural network models (i.e., ANN, LSTM-ANN, BLSTM-ANN,

GARCH-ANN, GARCH-LSTM-ANN, and GARCH-BLSTM-ANN) will be investigated in

this study. A comparative study on the proposed models will be conducted to find the best

approach for the prediction of copper price volatility.

3
 Journal Pre-proof

The remainder of this paper is structured as follows. In section 2, we introduce six

prediction models. In section 3, we conduct an empirical analysis to assess the prediction results.

In section 4, we summarize the main results and conclude this study.

of
2. Methodology

2.1 Data preprocessing method

Return estimation

pro
The closing price of copper has a very crucial role in the metal market. Moreover, investors are

more concerned with return rather than the price. Therefore, the volatility of returns is

forecasted in this paper. The returns of copper on day t are computed by using the following

equation: re-
Rt  (log pt  log pt 1 ) 100% (1)

where Rt is the return of the time series on day t , Pt and Pt 1 are the closing prices of

t t 1 , respectively.
lP
the financial time series on day and

Volatility estimation

Volatility plays a very important role in the metal market. As a classic metric of volatility,

variance reflects how much copper has varied during a certain period. In this paper, we compare
rna

the predicted and realized volatilities (i.e., variances) to assess the performance of various

models. Realized volatility ( RVt ) on day t during T trading days is calculated by the

following equation:
j t T 1
1

Jou

RVt  ( R j  R )2
T j t
(2)

where T is the number of trading days after day t , R j is the return of copper on day j ,

and R is the average return of copper during T trading days. RVt is the actual volatility.

In this study, we will consider three cases, i.e., T  10,15, 20 . For T  10 , RVt is the realized

volatility over the next 10 trading days (two weeks) from day t . Similarly, RVt values for

4
 Journal Pre-proof

T  15 and T  20 then correspond to the realized volatilities over the next 15 trading days

(three weeks) and 20 trading days (four weeks) from day t , respectively.

Data standardization

Standardization is very important before putting data into the ANN and LSTM networks. It can

of
enhance the training efficiency of the prediction models. In this paper, we choose the MinâMax

method to normalize the raw input data. Following this method, the input variables are

pro
calculated as follows:

xi ,t  xi ,min
xi ,t  (3)
xi ,max  xi ,min

Where xi ,t  [0,1] is the standardized value of the i th feature on day t , xi ,t is the actual
re-
value of the i th feature on day t , xi ,min and xi ,max is the minimum and maximum value of

the i th feature, respectively.

Measures of prediction errors

lP
To evaluate and compare the performances of the prediction models, four common measures

of prediction errors are used, including the mean squared error (MSE), the mean absolute error

(MAE), the mean absolute percentage error (MAPE), and the root mean squared error (RMSE).
rna

They are defined as follows:

N
1
MSE 
N
 ( PV  RV )
i 1
t t
2

(4)

1 N
MAE   PVt  RVt
N i 1 (5)
Jou

N
1
MAPE 
N
 1  PV / RV
i 1
t t
(6)

N
1
RMSE 
N
 ( PV  RV )
i 1
t t
2

(7)

where PVt and RVt are the predicted and realized volatilities of copper return series,

respectively; N is the number of predictions. A lower value of the measure indicates a better

5
 Journal Pre-proof

prediction.

We note that the four measures of prediction errors (i.e., MSE, MAE, RMSE, and MAPE)

introduced above are widely-applied in the literature. For example, Fuertes et al. (2009) and

Kim and Won (2018) have applied MSE, MAE, and MAPE in analyzing stock volatility

of
forecasts. Kristjanpoller et al. (2014) has utilized the four measures (MSE, MAE, RMSE, and

MAPE) to examine the volatility prediction results of hybrid neural network models for three

Latin-American stock exchange indexes. Other applications of these four measures in analyzing

pro
prediction models of commodity market include Bentes (2015), Zhang et al. (2015), and

Kristjanpoller and Hernández (2017). Besides, these four measures have also been adopted in

studying prediction models in the foreign exchange market (Sermpinis et al., 2012; Petropoulos

et al., 2017; Henriquez and Kristjanpoller 2019). To be in line with the literature, all the four
re-
measures of prediction errors (i.e., MSE, MAE, RMSE, and MAPE) are utilized in this study

to compare different volatility prediction models of copper price.

2.2 Prediction models

GARCH model
lP

GARCH model describes the variance of the current error term as a function of the error terms

of the previous periods. The specification for the GARCH(p, q) is defined as:

Rt    at (8)
rna

at   t t (9)

p q
 t2   0    i at2i    j t2 j (10)
i 1 j 1

Where  0  0 , i  0, i  1,..., p ,  j  0, j  1,..., q which guarantee that the conditional

Jou

p q
variance of GARCH(p, q) is always positive.   i    j  1 which ensures that the
i 1 j 1

variance is finite.

ANN model

ANN is a network of artificial neurons, which can receive inputs, change their internal states

according to the inputs, and then compute outputs based on the inputs and internal states. These

artificial neurons have weights that can be modified by a process called learning. Figure 1
6
 Journal Pre-proof

shows a neural network model that sequentially calculates the value of the output layer from

the input layer by using the output of the previous layer as the input for the current layer.

of
pro
Input Layer Hidden Layer Output Layer

Figure 1: An illustrative graph of the ANN architecture.

The ANN follows the following formulation:

y  f ( f ( xi wij  b j ) w jk  bk ) (11)
re-
Equation (11) shows that the input variable xi is multiplied by weight wij and summed

with bias b j , f (.) is the activation function, the result of this layer is the inputs of the next

layer, and the final result y is the forecast value.

lP
LSTM model

LSTM is a classic type of RNNs which can deal with the exploding and vanishing gradient

problems. It is normally augmented by recurrent gates called forget gates. Different from
rna

previous neural networks such as ANN, LSTM can learn very deep learning tasks that require

a long-time memory of events. LSTM can also handle inputs or signals that have both low and

high-frequency components. For more technical details, please see Hochreiter and

Schmidhuber (1997) and Gers et al. (2000). The structure of an LSTM unit is demonstrated in

Figure 2.
Jou

ht-1 xt ht-1 xt

It Input Gate Ot Output Gate

xt
ht-1
X Ct X ht
multiply multiply
X multiply

Ft Forget Gate

ht-1 xt

7
 Journal Pre-proof

Figure 2: An illustrative graph of the LSTM unit.

The equations for an LSTM unit are demonstrated in (12)-(16):

ft   (W fx xt  W fh ht 1  b f )
(12)

it   (Wix xt  Wih ht 1  bi ) (13)

of
ot   (Wox xt  Woh ht 1  bo ) (14)

ct  ft ct 1  it tanh(Wcx xt  Wch ht 1  bc ) (15)

pro
ht  ot tanh(ct ) (16)

As Figure 2 illustrates, the LSTM contains a memory cell ( ct ) and three gates: an input gate

( it ), a forget gate ( f t ), and an output gate ( ot ). In equations (12)-(16), the initial values are
re-
c0  0 and h0  0 , the operator denotes the element-wise product (i.e., Hadamard

product). At time t, xt represents the input vector, ht is the hidden state vector which is
lP
also known as the output vector of the LSTM unit. W and b are weight matrices and bias

parameters which need to be learned during training.  (.) is the sigmoid function and

tanh(.) is hyperbolic tangent function.

rna

Bidirectional LSTM model

The design of BLSTM is to access both the past and future information by combining a forward

hidden layer and a backward hidden layer (Liu and Guo, 2019). As shown in figure 3, BLSTM

can access long-range information in two opposite directions: one processing the sequential
Jou

data from top to bottom, the other one from bottom to top. The equations for a BLSTM unit are

given in (17)-(19):

ht  f (Wxh xt  Whh ht 1  bh )
(17)

ht  f (Wxh xt  Whh ht 1  bh )
(18)

yt  g (Why ht  Why ht  by ) (19)

where xt represents the input vector, ht is the hidden state vector. W and b are

8
 Journal Pre-proof

weight matrices and bias parameters which need to be learned during training. Figure 3 shows

the structure of the BLSTM unit.

xt 1 ht 1 ht 1 yt 1

of
xt ht ht yt

xt 1 ht 1 yt 1

pro
ht 1

Inputs Forward Layer Backward Layer Output

Figure 3: An illustrative graph of the BLSTM architecture.

LSTM-ANN and BLSTM-ANN networks

re-
LSTM-ANN and BLSTM-ANN are hybrid network models that combine LSTM and BLSTM

with ANN model. A simple illustration of the LSTM-ANN (BLSTM-ANN) structure is shown

in Figure 4.

Input LSTM block ANN block Output Input BLSTM block ANN block Output
lP

Panel A. LSTM-ANN Panel B. BLSTM-ANN

Figure 4: An illustrative graph of the hybrid neural network.

Six hybrid prediction models

rna

We now proceed to formally introduce the six prediction models of copper price volatility that

will be tested and analyzed in detail. By integrating the aforementioned LSTM-ANN and

BLSTM-ANN networks with the GARCH model, we can construct two hybrid prediction

models, namely, GARCH-LSTM-ANN and GARCH-BLSTM-ANN; they use the GARCH

Jou

forecasts as extra inputs in an attempt to provide useful complementary information to the

neural network model. The effectiveness of incorporating GARCH forecasts is then assessed

by benchmarking these two hybrid models with models of the same network structure but

without using the GARCH forecasts (i.e., LSTM-ANN and BLSTM-ANN). On the other hand,

to test the effectiveness of introducing the memory units (LSTM or BLSTM) into the hybrid

model, the memory-free ANN structure is also introduced for comparison purpose, resulting in

two additional benchmark models (ANN and GARCH-ANN). Thus, we have six different types

of hybrid prediction models, namely, ANN, LSTM-ANN, BLSTM-ANN, GARCH-ANN,

9
 Journal Pre-proof

GARCH-LSTM-ANN, and GARCH-BLSTM-ANN. We will use real-world data to compare

and test their prediction performances from different angles to develop insights on the

development of effective hybrid prediction models.

of
3. Empirical analysis

3.1. Data description

We collect data from January 1, 2008 to December 31, 2018, totaling eleven years. The data

pro
source is the WIND database. A group of explanatory variables are utilized, which could

potentially improve the volatility forecasts. The detailed description of the explanatory

variables is given in Table 1. The variables are related to the main metal prices, the main stock

market indices, currency, the main metal futures, and interest rate.
re-
Table 1: Variables and descriptions.

Group Variable Description

Metal markets CP Copper spot price of Yangtze River nonferrous metals, China

AP Aluminum spot price of Yangtze River nonferrous metals, China

lP

ZP Zinc spot price of Yangtze River nonferrous metals, China

GP Gold spot price of Shanghai nonferrous metals, China

LCP Copper spot price of London Metals Exchange

rna

LAP Aluminum spot price of London Metals Exchange

LZP Zinc spot price of London Metals Exchange

Exchange rates USD CNY - US Dollar exchange rate

EUR CNY - Euro exchange rate

Jou

YEN CNY - 100JPY exchange rate

Stock markets DJIA DJIA index of American stock market

FTSE FTSE 100 index of British stock market

CSI CSI300 index of Chinese stock market

Futures markets LMCP Copper futures price of London Metals Exchange

LMAP Aluminum futures price of London Metals Exchange

LMZP Zinc futures price of London Metals Exchange

10
 Journal Pre-proof

OIL WTI crude oil futures price

HCP Copper futures price of Shanghai Futures Exchange

HAP Aluminum futures price of Shanghai Futures Exchange

HZP Zinc futures price of Shanghai Futures Exchange

of
Interest rate SHI Shanghai Interbank Offered Rate

We use the daily closing prices of these variables to compute their daily returns. Missing

values are filled with the last valid values of previous trading days. Then, each dataset is

pro
sequentially divided into two sets, i.e., the training set which contains 70% of the data, and the

testing set which contains the remaining 30% data. Table 2 presents the common statistics for

returns of the variables.

Table 2: Descriptive statistics for the returns of the variables.

re-
Variable Mean (%) Std. Dev.(%) Min (%) Max (%) Jarque-Bera ADF ARCH（12）
CP -0.0075 1.33 -9.06 8.05 6823.15 -22.23 634.97
AP -0.0103 0.85 -6.01 4.29 4389.83 -44.49 505.86
ZP 0.0049 1.29 -7.04 7.27 1809.67 -47.59 303.92
GP 0.0137 0.96 -7.35 6.55 7295.08 -43.90 219.09
lP
LCP -0.0038 1.72 -19.47 11.37 12323.63 -53.89 397.11
LAP -0.0086 1.49 -11.00 8.50 2294.83 -53.88 153.24
LZP 0.0024 1.89 -15.46 9.80 1695.34 -52.65 328.49
USD -0.0023 0.15 -0.93 1.84 41083.53 -46.70 680.29
EUR -0.0115 0.62 -6.93 3.34 7425.28 -51.18 124.67
rna

YEN -0.0020 0.64 -5.15 3.63 3387.94 -38.22 182.25

DJIA 0.0213 1.20 -11.27 10.51 17587.75 -40.42 792.44
FTSE 0.0018 1.22 -10.33 9.38 10556.78 -24.65 633.30
CSI -0.0217 1.75 -9.15 8.93 1807.47 -50.32 334.43
LMCP -0.0045 1.73 -19.76 11.88 12873.50 -23.71 383.21
LMAP -0.0103 1.39 -10.56 7.54 1780.60 -53.79 212.34
LMZP 0.0001 1.88 -15.27 9.78 1680.29 -54.27 331.47
Jou

OIL -0.0294 2.51 -19.66 16.41 4000.99 -55.11 483.10

HCP -0.0071 1.39 -7.50 6.71 1531.31 -34.52 453.46
HAP -0.0100 0.99 -6.94 5.40 4616.02 -52.67 450.73
HZP 0.0041 1.53 -9.40 5.96 1149.04 -54.66 251.10
SHI -0.0210 7.70 -62.48 77.52 63021.49 -36.95 228.08

Note: The critical value at 5% for Jarque–Bera test is 5.99. ADF is the stationarity test, and the critical

value at 5% is −2.86. The ARCH(12) statistic corresponds to the ARCH-LM test with 12 lags, where the

probability distribution is  2 (12) and the critical value at the 5% level is 21.03.

11
 Journal Pre-proof

From Table 2, it is shown that the mean copper returns (CP) is close to 0 (-0.0075%), and

its standard deviation is close to 1%. The ADF test is significant at 5%, which indicates that the

time series are stationary. However, the Jarque–Bera (normality) test shows that the copper

returns are not normally distributed. Besides, ARCH (12) is the heteroskedasticity test used to

of
identify the presence of ARCH effects for copper returns, and the null hypothesis is rejected,

indicating the existence of ARCH effects. Moreover, Table 2 shows that other explanatory

variables also exhibit some complex statistical features, such as non-normality and ARCH

pro
effects. These complex statistical features imply that it is difficult to deal with these time series

with conventional statistical methods. It is therefore meaningful to investigate the effective

combination of both conventional statistical methods (e.g., GARCH) and the state-of-art deep

learning techniques (e.g., LSTM) to improve the copper price volatility prediction.
re-
Next, we depict the copper price, as well as its variance of three weeks from 2008 to 2018

in Figure 5.
lP
rna

Figure 5: Time series plots for daily closing price and the variance of three weeks.

From Figure 5, it can be observed that, in the second half of 2008, the price of copper

suddenly dived, coinciding with the subprime crisis. This unexpected crash greatly increased
Jou

the subsequent market volatility in the second half of 2008 and the beginning of 2009.

Furthermore, high volatility is also observed in 2011, 2013, and 2016.

The lower the correlation between the explanatory variables indicating more information

can be taken by the models to make a better fit. For this reason, the study on explanatory

variables is conducted with the correlation analysis and principal component analysis. Both

analyses are investigated in terms of returns. The heat map of return correlation coefficient

matrix is shown in Figure 6.

12
 Journal Pre-proof

of
pro
Figure 6: Heat map of return correlation coefficient matrix.

From Figure 6, we can observe that the correlation coefficients disperse. It means that we

need some methods to extract useful information. Also, from the principal component analysis
re-
(shown in Appendix) we can find that with 19 eigenvectors, 98.85% of explained variance ratio

is achieved. Furthermore, 99.45% is achieved with 20 eigenvectors. These results also show

that the relationship is not high.

lP

3.2. Empirical results

In this section, six different types of hybrid prediction models are empirically tested and

compared, including ANN, LSTM-ANN, BLSTM-ANN, GARCH-ANN, GARCH-LSTM-

rna

ANN, and GARCH-BLSTM-ANN. The empirical study aims to unravel four main questions:

1) Is the inclusion of GARCH forecasts as network inputs an effective way to enhance the

prediction performance? 2) Can the integration of RNNs with GARCH forecasts significantly

reduce the prediction errors? 3) Which neural network architecture, particularly the choice
Jou

between LSTM and BLSTM, works better for the task of copper price volatility prediction? 4)

How to fine-tune the ANN configuration (e.g., the number of layers and neurons) in the hybrid

network to optimize the prediction results?

To this end, different choices of the number of hidden layers and neurons in each layer are

tested and compared. Each ANN configuration is denoted by ANN (l,n), where l corresponds

to the number of hidden layers, and n corresponds to the number of neurons in each layer. We

set the hidden layers and neurons of ANN according to the results of Kristjanpoller and

13
 Journal Pre-proof

Hernández (2017); thus, l  4,5, 6 and n  10, 20 . As for the GARCH model, we have

tested the model performances of various parameters values ( p and q , see equation (10))

using the copper price data, and found that in general the best choice is p  1 and q  1 . Thus,

GARCH(1,1) is adopted in the numerical experiments.

of
In the first set of experiments, we use one year trading data (i.e., one-year input window,

including 252 trading days) to forecast the volatility in the next two weeks (two-week ahead

pro
forecasts, or T  10 in equation (2)). Four measures of prediction errors (i.e., MSE, MAPE,
MAE, and RMSE) are utilized to evaluate the prediction performances. To facilitate the

prediction error analysis, all returns are scaled up by a factor of 100. The results are presented

in Table 3. re-
Table 3: Prediction performances of two-week ahead forecasts for different hybrid neural

network configurations. (Var% is the percentage variation of the error measures. S1, S2 and S3

represent the ANN, LSTM-ANN and BLSTM-ANN respectively)

lP

Panel A

Measure of prediction errors MSE MAPE

Network architecture Ranking (l,n) MSE Var% (l,n) MAPE Var%

rna

GARCH 1.03E-02 -253.95% 14.7829 -311.53%

S1 1 (6,10) 4.26E-03 -46.39% (6,10) 5.8649 -63.27%

2 (6,20) 4.60E-03 -58.08% (6,20) 7.4989 -108.76%

3 (4,20) 4.93E-03 -69.42% (4,20) 8.3441 -132.28%

Jou

S2 1 (6,20) 4.35E-03 -49.48% (6,10) 4.4011 -22.52%

2 (6,10) 4.41E-03 -51.55% (6,20) 6.5909 -83.48%

3 (4,20) 4.60E-03 -58.08% (4,10) 7.474 -108.06%

S3 1 (6,10) 4.38E-03 -50.52% (6,10) 6.6012 -83.76%

2 (4,20) 4.56E-03 -56.70% (6,20) 7.3599 -104.89%

3 (6,20) 4.57E-03 -57.04% (4,20) 7.3894 -105.71%

GARCH-S1 1 (4,10) 4.53E-03 -55.67% (6,10) 5.6488 -57.25%

14
 Journal Pre-proof

2 (5,10) 4.66E-03 -60.14% (5,10) 5.9565 -65.82%

3 (5,20) 4.68E-03 -60.82% (5,20) 7.722 -114.97%

GARCH-S2 1 (6,10) 2.91E-03 base case (5,20) 3.5922 base case

2 (5,10) 3.55E-03 -21.99% (6,20) 3.8292 -6.60%

of
3 (6,20) 3.56E-03 -22.34% (4,20) 4.2919 -19.48%

GARCH-S3 1 (6,20) 3.53E-03 -21.31% (6,10) 4.264 -18.70%

2 (5,10) 3.56E-03 -22.34% (6,20) 4.3101 -19.98%

pro
3 (5,20) 3.63E-03 -24.74% (5,10) 4.6163 -28.51%

Panel B

Measure of prediction errors MAE RMSE

Network architecture Ranking (l,n) MAE

re- Var% (l,n) RMSE Var%

GARCH 9.26E-02 -192.11% 1.01E-01 -87.38%

S1 1 (6,10) 3.72E-02 -17.35% (6,10) 6.53E-02 -21.15%

2 (6,20) 4.53E-02 -42.90% (6,20) 6.78E-02 -25.79%

lP
3 (4,20) 5.04E-02 -58.99% (4,20) 7.02E-02 -30.24%

S2 1 (6,10) 3.15E-02 0.63% (6,20) 6.60E-02 -22.45%

2 (6,20) 4.06E-02 -28.08% (6,10) 6.64E-02 -23.19%

3 (4,20) 4.52E-02 -42.59% (4,20) 6.78E-02 -25.79%

rna

S3 1 (6,10) 4.05E-02 -27.76% (6,10) 6.62E-02 -22.82%

2 (4,20) 4.44E-02 -40.06% (4,20) 6.75E-02 -25.23%

3 (6,20) 4.45E-02 -40.38% (6,20) 6.76E-02 -25.42%

GARCH-S1 1 (4,10) 3.86E-02 -21.77% (4,10) 6.73E-02 -24.86%

Jou

2 (6,10) 3.92E-02 -23.66% (5,10) 6.82E-02 -26.53%

3 (5,10) 4.39E-02 -38.49% (5,20) 6.84E-02 -26.90%

GARCH-S2 1 (6,20) 3.17E-02 base case (6,10) 5.39E-02 base case

2 (5,20) 3.35E-02 -5.68% (5,10) 5.96E-02 -10.58%

3 (4,20) 3.36E-02 -5.99% (6,20) 5.96E-02 -10.58%

GARCH-S3 1 (6,20) 3.20E-02 -0.95% (6,20) 5.94E-02 -10.20%

2 (6,10) 3.38E-02 -6.62% (5,10) 5.97E-02 -10.76%

15
 Journal Pre-proof

3 (5,10) 3.43E-02 -8.20% (5,20) 6.02E-02 -11.69%

From Table 3, we have the following key observations. First, although the network models

can outperform the GARCH model, incorporating GARCH forecasts as inputs can enhance the

prediction power of the network models. For example, the MSE of GARCH-LSTM-ANN(6,20)

of
is 3.56E-03, which is smaller than LSTM-ANN(6,20) with a MSE of 4.35E-03. The MAPE of

GARCH-LSTM-ANN (6,10) is 4.2640, being smaller than that of LSTM-ANN(6,10), of which

the MAPE is 4.4011. Similar results can be found in the cases of MAE and RMSE in spite of a

pro
few exceptions. Second, both LSTM and BLSTM can improve the prediction performance of

memory-free ANN. In other words, both GARCH-LSTM-ANN and GARCH-BLSTM-ANN

models outperform the GARCH-ANN model, and both LSTM-ANN and BLSTM-ANN

models demonstrate better performances than the ANN model, with only few exceptions.
re-
Next, GARCH-LSTM-ANN architecture is better than the other five network architectures.

The MSE, MAPE, MAE, and RMSE of GARCH-LSTM-ANN models are the lowest in general.

There is only one exception. The MAE of LSTM-ANN(6,10) and GARCH-LSTM-ANN are

almost the same. Then, let us focus on GARCH-LSTM-ANN models. In terms of MSE, the
lP

best prediction model is GARCH-LSTM-ANN(6,10) with a lowest MSE of 2.91E-03. However,

the smallest MAPE (3.5922) is achieved by GARCH-LSTM-ANN(5,20). Besides, the

GARCH-LSTM-ANN architecture with the 6-layer and 20-neuron in ANN configuration is the
rna

best prediction model in terms of MAE. These results suggest that the ANN configuration

should be optimized depending on the choice of the measure of prediction errors.

We next investigate the case of three-week ahead forecasts ( T  15 in equation (2)). The

results are presented in Table 4.

Jou

Table 4: Prediction performances of three-week ahead forecasts for different hybrid neural

network configurations. (Var% is the percentage variation of the error measures. S1, S2 and S3

represent the ANN, LSTM-ANN and BLSTM-ANN respectively)

Panel A

Measure of prediction errors MSE MAPE

Network architecture Ranking (l,n) MSE Var% (l,n) MAPE Var%

16
 Journal Pre-proof

GARCH 1.10E-02 -237.42% 5.1202 -413.82%

S1 1 (5,20) 4.26E-03 -30.67% (5,20) 2.3503 -135.86%

2 (4,20) 4.67E-03 -43.25% (4,20) 2.6689 -167.83%

3 (4,10) 4.68E-03 -43.56% (4,10) 2.6963 -170.58%

of
S2 1 (6,10) 3.81E-03 -16.87% (5,20) 1.7711 -77.73%

2 (5,20) 3.85E-03 -18.10% (6,10) 1.9936 -100.06%

3 (6,20) 4.03E-03 -23.62% (6,20) 2.0984 -110.58%

pro
S3 1 (6,10) 3.87E-03 -18.71% (6,10) 1.8797 -88.63%

2 (5,20) 3.90E-03 -19.63% (5,20) 1.8994 -90.61%

3 (4,20) 4.05E-03 -24.23% (4,10) 2.0935 -110.09%

GARCH-S1 1 (6,20) 4.51E-03

re- -38.34% (5,20) 1.1374 -14.14%

2 (5,20) 4.65E-03 -42.64% (6,20) 2.5523 -156.13%

3 (5,10) 4.71E-03 -44.48% (5,10) 2.7035 -171.30%

GARCH-S2 1 (5,20) 3.36E-03 -3.07% (5,20) 1.1095 -11.34%

2 (5,10) 3.88E-03 -19.02% (4,10) 1.8645 -87.10%

lP
3 (4,10) 3.89E-03 -19.33% (5,10) 1.867 -87.36%

GARCH-S3 1 (4,10) 3.26E-03 base case (6,10) 0.9965 base case

2 (5,10) 3.46E-03 -6.13% (4,20) 1.0071 -1.06%

rna

3 (4,20) 3.60E-03 -10.43% (4,10) 1.0619 -6.56%

Panel B

Measure of prediction error MAE RMSE

Network architecture Ranking (l,n) MAE Var% (l,n) RMSE Var%

Jou

GARCH 9.83E-02 -248.58% 1.05E-01 -83.89%

S1 1 (5,20) 4.63E-02 -64.18% (5,20) 6.52E-02 -14.19%

2 (4,20) 5.18E-02 -83.69% (4,20) 6.83E-02 -19.61%

3 (4,10) 5.22E-02 -85.11% (4,10) 6.84E-02 -19.79%

S2 1 (5,20) 3.77E-02 -33.69% (6,10) 6.17E-02 -8.06%

2 (6,10) 4.09E-02 -45.04% (5,20) 6.21E-02 -8.76%

3 (6,20) 4.25E-02 -50.71% (6,20) 6.35E-02 -11.21%

17
 Journal Pre-proof

S3 1 (6,10) 3.91E-02 -38.65% (6,10) 6.22E-02 -8.93%

2 (5,20) 3.95E-02 -40.07% (5,20) 6.25E-02 -9.46%

3 (4,10) 4.24E-02 -50.35% (4,20) 6.36E-02 -11.38%

GARCH-S1 1 (5,20) 3.68E-02 -30.50% (5,20) 6.71E-02 -17.51%

of
2 (6,20) 4.97E-02 -76.24% (6,20) 6.82E-02 -19.44%

3 (5,10) 5.23E-02 -85.46% (5,10) 6.86E-02 -20.14%

GARCH-S2 1 (5,20) 2.81E-02 0.35% (5,20) 5.79E-02 -1.40%

pro
2 (5,10) 3.90E-02 -38.30% (5,10) 6.23E-02 -9.11%

3 (4,10) 3.99E-02 -41.49% (4,10) 6.23E-02 -9.11%

GARCH-S3 1 (4,20) 2.82E-02 base case (4,10) 5.71E-02 base case

2 (6,10) 2.92E-02
re- -3.55% (5,10) 5.88E-02 -2.98%

3 (4,10) 3.01E-02 -6.74% (4,20) 6.00E-02 -5.08%

From Table 4, we have the following key observations. First, we can find that the GARCH

model underperforms the network models and incorporating GARCH forecasts as inputs can

still enhance the prediction power of the network models in general. This can be illustrated by
lP

the following comparisons. Let us compare LSTM-ANN and GARCH-LSTM-ANN models.

The MSE of GARCH-LSTM-ANN(5,20) is 3.36E-03, which is smaller than that of LSTM-

ANN models. As for the comparison between BLSTM-ANN and GARCH-BLSTM-ANN, the
rna

MSE of GARCH-BLSTM-ANN(4,20) is 3.60E-03, which is significantly smaller than that of

BLSTM-ANN models. Similar results can be found for other error measures (MAPE, MAE

and RMSE) and the comparison between ANN and GARCH-ANN models. Besides, GARCH-

BLSTM-ANN architecture appears to be better than the other five network architectures, since
Jou

the MSE, MAPE, MAE, and RMSE of GARCH-BLSTM-ANN models are always the lowest.

We next examine the best network structure. It is clear that both LSTM-ANN and BLSTM-

ANN models demonstrate better performances than ANN models. Thus, the incorporation of

LSTM or BLSTM is effective. As for the ANN configurations (hidden layers and neurons), it

still appears that the best configuration should be dependent on the choice of error measures.

After the investigation of two-week and three-week ahead forecasts, we then turn to four-

week ahead forecasts ( T  20 in equation (2)) to analyze the robustness of the results. The

results are summarized in Table 5.

18
 Journal Pre-proof

Table 5: Prediction performances of four-week ahead forecasts for different hybrid neural

network configurations. (Note: Var% is the percentage variation of the error measures. S1, S2

and S3 represent the ANN, LSTM-ANN and BLSTM-ANN respectively)

of
Panel A

Measure of prediction errors MSE MAPE

Network architecture Ranking (l,n) MSE Var% (l,n) MAPE Var%

pro
GARCH 1.35E-02 -258.09% 4.8537 -639.89%

S1 1 (5,10) 4.19E-03 -11.14% (5,10) 1.6659 -153.95%

2 (4,20) 4.71E-03 -24.93% (4,20) 2.1890 -233.69%

3 (4,10) 4.92E-03
re- -30.50% (4,10) 2.2993 -250.50%

S2 1 (4,10) 4.05E-03 -7.43% (4,10) 1.4334 -118.51%

2 (5,20) 4.22E-03 -11.94% (5,20) 1.6980 -158.84%

3 (4,20) 4.36E-03 -15.65% (5,10) 1.7951 -173.64%

lP
S3 1 (5,20) 4.08E-03 -8.22% (4,10) 1.5288 -133.05%

2 (4,10) 4.16E-03 -10.34% (5,20) 1.6633 -153.55%

3 (5,10) 4.27E-03 -13.26% (5,10) 1.7930 -173.32%

GARCH-S1 1 (5,20) 4.62E-03 -22.55% (5,20) 1.8373 -180.08%

rna

2 (6,10) 4.70E-03 -24.67% (6,10) 2.1475 -227.36%

3 (4,20) 4.95E-03 -31.30% (6,20) 2.2763 -247.00%

GARCH-S2 1 (6,10) 3.77E-03 base case (4,20) 0.656 base case

2 (5,10) 3.87E-03 -2.65% (5,10) 0.8552 -30.37%

Jou

3 (6,20) 3.90E-03 -3.45% (6,10) 0.9237 -40.81%

GARCH-S3 1 (5,20) 3.65E-03 3.18% (6,20) 1.1977 -82.58%

2 (5,10) 4.16E-03 -10.34% (4,20) 1.2629 -92.52%

3 (6,20) 4.17E-03 -10.61% (5,10) 1.3316 -102.99%

Panel B

Measure of prediction errors MAE RMSE

Network architecture Ranking (l,n) MAE Var% (l,n) RMSE Var%

19
 Journal Pre-proof

GARCH 1.10E-01 -280.62% 1.16E-01 -88.93%

S1 1 (5,10) 4.12E-02 -42.56% (5,10) 6.47E-02 -5.37%

2 (4,20) 5.04E-02 -74.39% (4,20) 6.86E-02 -11.73%

3 (4,10) 5.30E-02 -83.39% (4,10) 7.01E-02 -14.17%

of
S2 1 (4,10) 3.77E-02 -30.45% (4,10) 6.36E-02 -3.58%

2 (5,20) 4.18E-02 -44.64% (5,20) 6.50E-02 -5.86%

3 (5,10) 4.35E-02 -50.52% (4,20) 6.60E-02 -7.49%

pro
S3 1 (4,10) 3.93E-02 -35.99% (5,20) 6.39E-02 -4.07%

2 (5,20) 4.10E-02 -41.87% (4,10) 6.45E-02 -5.05%

3 (5,10) 4.31E-02 -49.13% (5,10) 6.54E-02 -6.51%

GARCH-S1 1 (5,20) 4.44E-02

re- -53.63% (5,20) 6.80E-02 -10.75%

2 (6,10) 4.99E-02 -72.66% (6,10) 6.85E-02 -11.56%

3 (6,20) 5.25E-02 -81.66% (4,20) 7.03E-02 -14.50%

GARCH-S2 1 (6,10) 2.89E-02 base case (6,10) 6.14E-02 base case

2 (5,10) 2.92E-02 -1.04% (5,10) 6.22E-02 -1.30%

lP
3 (4,20) 2.96E-02 -2.42% (6,20) 6.25E-02 -1.79%

GARCH-S3 1 (6,20) 3.50E-02 -21.11% (5,20) 6.04E-02 1.63%

2 (5,10) 3.64E-02 -25.95% (5,10) 6.45E-02 -5.05%

rna

3 (5,20) 3.69E-02 -27.68% (6,20) 6.45E-02 -5.05%

From Table 5, we have the following key observations. First, we can find that

incorporating GARCH forecasts as inputs can enhance the prediction power of the network

models in general. For example, the MSE of GARCH-BLSTM-ANN(5,20) is 3.65E-03, being

Jou

smaller than that of BLSTM-ANN(5,20), of which the MSE is 4.08E-03. Besides, it is clear

that both LSTM and BLSTM can improve the prediction performance of ANN, which is

consistent with former observations for the cases of the two-week and three-week forecasts.

Next, GARCH-LSTM-ANN architecture still presents the best predictive power than the

other five network architectures in general. For example, the MAPE of GARCH-LSTM-

ANN(6,10) is 0.9237; while this is the worst among all GARCH-LSTM-ANN model

configurations, it is still better than other prediction models. Then, let us examine the influence

of network configuration to optimize the prediction results. We find that when prediction errors
20
 Journal Pre-proof

are measured in terms of MSE, MAPE, MAE, and RMSE, the best prediction models are

GARCH-LSTM-ANN(6,10), GARCH-LSTM-ANN(4,20), GARCH-LSTM-ANN(6,10) and

GARCH-LSTM-ANN(6,10), respectively. Thus, the best configurations (hidden layers and

neurons) of the prediction model are a bit different for different error measures.

of
To sum up, we can find the following conclusions from Tables 3, 4 and 5. First,

incorporating GARCH forecasts as inputs can enhance the prediction power of the network

models in general. Second, the integration of memory networks (i.e., LSTM and BLSTM) with

pro
the classical ANN can improve the prediction power of the entire neural network. Third, the

best model architecture for volatility prediction is GARCH-LSTM-ANN for two-week and

four-week ahead forecasts. However, for three-week ahead forecasts, the best architecture is

GARCH-BLSTM-ANN. Fourth, the ANN configuration in the hybrid model should be

re-
optimized depending on the choice of the measure of prediction errors.

Further, the predicted and realized copper price volatilities in both the training and testing

sets are shown in the following Figure 7. The cases of two-week, three-week, and four-week

ahead forecasts are all considered. Figure 7 shows that the deviation between the predicted and
lP

realized volatilities will generally become larger as the forecast horizon grows from two weeks

to four weeks. This is consistent with the intuition that it should be easier to make predictions

about the nearer future. Moreover, from Figure 7, it seems that the forecasting accuracy is not
rna

dependent on volatility level. For instance, for the two-week and three-week ahead forecasts in

the testing set, the realized volatilities around November 2016 are extremely high, whereas the

realized volatilities around August 2017 are relatively low. However, there are significantly

large prediction errors around both November 2016 and August 2017. Thus, the size of the
Jou

prediction error is not necessarily dependent on the level of the realized volatility.

21
 Journal Pre-proof

of
pro
re-
lP
rna

Figure 7: Predicted and realized copper price volatilities

Jou

4. Conclusions

We have investigated the integration of deep learning methodology and GARCH model to

improve the prediction of copper price volatility. Facing the long memory phenomenon in time

series data, recurrent neural networks (RNNs) are inherently suitable to distill information from

sequences of inputs. Thus, the conventional ANN is combined with two RNNs (LSTM and

BLSTM) to generate hybrid neural networks. Besides, GARCH models are trained and their

best forecasts are used as extra inputs to augment the training data for the hybrid neural

networks. As a result, six different types of hybrid neural network models (i.e., ANN, LSTM-
22
 Journal Pre-proof

ANN, BLSTM-ANN, GARCH-ANN, GARCH-LSTM-ANN, and GARCH-BLSTM-ANN)

are developed and tested. We further conduct a series of sensitivity analysis of the ANN

configuration (i.e., hidden layers and neurons) for each neural network architecture to optimize

the prediction performance.

of
Through the empirical analyses, we have several major findings that can contribute to the

literature. First, we find the GARCH forecasts can serve as informative features to substantially

boost the volatility prediction, which complements the results of Kristjanpoller and Hernández

pro
(2017). We also find that incorporating RNNs (LSTM and BLSTM) into the hybrid GARCH–

ANN network can further improve the volatility prediction performance; such a finding

highlights the efficacy of deep learning methodology and advances the results of Kristjanpoller

and Hernández (2017). Besides, we have conducted a variety of experiments to find the best
re-
architecture of the hybrid neural network. In particular, we find that the choice between LSTM

and BLSTM networks should be dependent on the forecast horizon. It appears that BLSTM

works better for three-week ahead volatility forecasts while LSTM works better for other cases.

Further, the empirical results suggest that the ANN configuration in the hybrid model should
lP

be optimized depending on the choice of the measure of prediction errors. From these findings,

an effective hybrid approach is developed to construct useful volatility prediction models

through neatly integrating GARCH forecasts with ANN, LSTM, and BLSTM networks. This
rna

proposed hybrid approach, we hope, will motivate further investigation into the exploration and

exploitation of using deep learning methodology in solving time series forecasting problems.

Appendix
Jou

Table A.1: Principal component analysis results. (Note: CuP represents cumulative proportion

of explained variance)

Panel A

Variable PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10

AP 0.21 -0.24 -0.14 -0.10 0.57 0.05 0.09 0.14 -0.13 0.08

CP 0.27 -0.27 -0.05 -0.23 -0.06 0.08 0.10 -0.20 -0.05 0.16

CSI 0.15 -0.01 -0.12 -0.05 -0.40 -0.04 -0.28 0.73 0.14 0.18

23
 Journal Pre-proof

DJIA 0.14 0.32 -0.01 -0.47 0.00 0.17 0.17 0.08 0.09 -0.39

EUR 0.13 -0.13 0.48 -0.20 0.01 -0.04 -0.11 0.01 0.19 0.34

FTSE 0.20 0.32 -0.04 -0.40 -0.03 0.07 0.07 0.04 0.15 -0.16

GP 0.09 -0.13 0.55 -0.08 -0.07 0.07 0.31 0.00 0.17 0.24

of
HAP 0.25 -0.16 -0.14 0.07 0.47 0.03 -0.02 0.37 0.02 -0.05

HCP 0.29 -0.23 -0.06 0.01 -0.17 0.07 0.01 0.04 0.05 0.01

HZP 0.29 -0.21 -0.07 0.11 -0.24 0.06 0.04 0.05 -0.20 -0.21

pro
LAP 0.26 0.02 0.09 0.23 0.24 -0.18 -0.12 -0.11 0.48 -0.23

LCP 0.32 0.00 0.04 0.09 -0.08 -0.04 -0.05 -0.17 0.27 -0.10

LMAP 0.22 0.34 0.09 0.23 0.20 -0.10 -0.06 -0.03 -0.05 0.22

LMCP 0.27 0.33 0.00 0.10 re- -0.02 0.01 0.01 -0.04 -0.17 0.24

LMZP 0.25 0.31 0.02 0.26 -0.10 -0.03 -0.01 -0.02 -0.35 0.16

LZP 0.30 -0.01 0.03 0.26 -0.20 -0.07 -0.05 -0.16 0.06 -0.31

OIL 0.17 0.30 0.01 -0.17 0.09 0.13 0.10 0.00 -0.15 0.19

SHI -0.01 0.00 0.11 0.07 0.07 0.81 -0.54 -0.13 0.01 -0.02
lP
USD -0.04 0.04 -0.36 0.31 -0.09 0.42 0.55 0.01 0.40 0.26

YEN -0.07 0.02 0.50 0.28 0.04 0.20 0.33 0.37 -0.22 -0.37

ZP 0.25 -0.30 -0.06 -0.11 -0.14 0.08 0.15 -0.18 -0.35 -0.03
rna

CuP (%) 30.39 42.42 49.41 54.77 59.76 64.62 69.16 73.39 77.12 80.67

Panel B

Variable PC11 PC12 PC13 PC14 PC15 PC16 PC17 PC18 PC19 PC20 PC21

AP 0.03 -0.14 0.05 0.19 0.28 -0.20 -0.08 0.29 0.46 0.05 -0.04
Jou

CP -0.12 -0.05 -0.32 0.31 -0.02 0.01 0.36 -0.19 -0.25 0.42 -0.30

CSI -0.09 -0.07 0.10 0.31 0.06 -0.07 -0.06 0.01 -0.02 0.01 -0.01

DJIA 0.17 -0.10 -0.01 0.05 -0.39 -0.47 -0.02 0.00 0.00 0.01 0.05

EUR 0.59 0.40 -0.08 -0.06 0.01 -0.05 -0.03 0.01 0.06 -0.02 0.00

FTSE 0.13 -0.16 0.01 -0.11 0.43 0.61 0.11 0.02 0.05 -0.04 -0.01

GP -0.33 -0.44 0.37 -0.19 -0.01 -0.07 -0.01 0.00 -0.02 -0.01 -0.01

HAP 0.04 0.03 0.05 -0.43 -0.05 0.01 0.07 -0.41 -0.41 -0.07 0.04

24
 Journal Pre-proof

HCP -0.21 0.05 -0.39 -0.26 -0.28 0.08 0.34 0.24 0.29 -0.44 0.14

HZP 0.14 0.08 0.23 -0.31 -0.28 0.24 -0.28 0.16 0.19 0.47 -0.20

LAP -0.06 0.08 0.16 0.27 -0.16 0.13 0.07 0.46 -0.31 0.02 -0.04

LCP -0.21 0.05 -0.32 -0.06 0.26 -0.15 -0.41 -0.21 0.08 0.25 0.48

of
LMAP 0.07 -0.22 -0.02 0.28 -0.44 0.28 -0.03 -0.39 0.33 -0.01 0.04

LMCP -0.03 -0.11 -0.31 -0.14 0.08 -0.12 -0.42 0.22 -0.25 -0.23 -0.48

LMZP 0.23 -0.14 0.09 -0.11 0.15 -0.20 0.43 0.25 -0.15 0.22 0.39

pro
LZP 0.08 0.08 0.26 0.04 0.30 -0.28 0.22 -0.34 0.23 -0.25 -0.37

OIL -0.47 0.67 0.29 0.03 -0.01 0.01 0.04 -0.01 0.05 0.02 0.03

SHI -0.02 -0.09 0.05 0.02 0.01 0.01 0.00 0.01 -0.01 0.00 0.00

USD 0.22 0.08 0.03 0.05 0.01

re- 0.00 -0.01 -0.01 0.01 0.00 0.01

YEN -0.02 0.14 -0.31 0.22 0.11 0.12 0.03 0.00 -0.03 0.03 -0.01

ZP 0.15 -0.02 0.22 0.36 -0.02 0.15 -0.27 -0.04 -0.28 -0.41 0.31

CuP (%) 83.88 86.87 89.58 91.99 93.95 95.71 96.94 97.92 98.85 99.45 100
lP

References

Azadeh, A., Moghaddam, M., Khakzad, M., & Ebrahimipour, V. (2012). A flexible neural

network-fuzzy mathematical programming algorithm for improvement of oil price

rna

estimation and forecasting. Computers and Industrial Engineering, 62(2), 421–430.

Bentes, S.R. (2015). Forecasting volatility in gold returns under the GARCH, IGARCH and

FIGARCH frameworks: New evidence. Physica A: Statistical Mechanics and its

Applications, 438, 355-364.

Jou

Bergerson, K., & Wunsch, D. C. (1991). A commodity trading model based on a neural

network-expert system hybrid. Proceedings of the International Joint Conference on

Neural Networks, 91, 289–293.

Bildirici, M., & Ersin, Ö. Ö. (2013). Forecasting oil prices: Smooth transition and neural

network augmented GARCH family models. Journal of Petroleum Science and

Engineering, 109, 230–240.

Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of

Econometrics, 31(3), 307–327.

25
 Journal Pre-proof

Cui, L., Huang, K., & Cai, H. J. (2015). Application of a TGARCH-wavelet neural network to

arbitrage trading in the metal futures market in China. Quantitative Finance, 15(2), 371–

384.

Davidson, J. (2004). Moment and memory properties of linear conditional heteroscedasticity

of
models, and a new model. Journal of Business & Economic Statistics, 1, 16–29.

Ding, Z., Engle, R., & Granger, C. (1993). A long memory property of stock market returns and

a new model. Journal of Empirical Finance, 1, 83–106.

pro
Eapen, J., Bein, D., & Verma, A. (2019). Novel deep learning model with CNN and bi-

directional LSTM for improved stock market index prediction. IEEE 9th Annual

Computing and Communication Workshop and Conference (CCWC), 0264-0270.

Ederington, L. H., & Lee, J. H. (1993). How Markets Process Information: News Releases and
re-
Volatility. Journal of Finance, 48(4), 1161-1191.

Fuertes, A. M., Izzeldin, M., & Kalotychou, E. (2009). On forecasting daily stock volatility: the

role of intraday information and market conditions. International Journal of Forecasting,

25(2), 259-281.
lP

Garcí
a, D., & Kristjanpoller, W. (2019). An adaptive forecasting approach for copper price

volatility through hybrid and non-hybrid models. Applied Soft Computing, 74, 466-478.

Graves, A., Liwicki, M., Fernández, S., Bertolami, R., Bunke, H., & Schmidhuber, J. (2009). A
rna

novel connectionist system for improved unconstrained handwriting recognition. IEEE

Transactions on Pattern Analysis and Machine Intelligence. 31 (5), 855–868.

Hamid, S. A., & Iqbal, Z. (2004). Using neural networks for forecasting volatility of S&P 500

Index futures prices. Journal of Business Research, 57(10), 1116–1125.

Jou

Hamilton, J. D., & Susmel, R. (1994). Autoregressive conditional heteroscedasticity and

changes in regime. Journal of Econometrics, 64, 307–333.

Henriquez, J., & Kristjanpoller, W. (2019). A combined Independent Component Analysis–

Neural Network model for forecasting exchange rate variation. Applied Soft Computing.

Herman, M., & Schrauwen, B. (2013). Training and analyzing deep recurrent neural networks.

Advances in neural information processing systems(NIPS). 26.

Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8),

1735–1780.
26
 Journal Pre-proof

Kim, H. Y. & Won C. H. (2018). Forecasting the volatility of stock price index: A hybrid model

integrating LSTM with multiple GARCH-type models. Expert Systems with Applications,

103, 25-37.

Kristjanpoller, W., Fadic, A., & Minutolo, M. C. (2014). Volatility forecast using hybrid neural

of
network models. Expert Systems with Applications, 41(5), 2437-2442.

Kristjanpoller, W., & Hernández P, E. (2017). Volatility of main metals forecasted by a hybrid

ANN-GARCH model with regressors. Expert Systems with Applications 84, 290-300.

pro
Kristjanpoller, W., & Minutolo, M. (2015). Gold price volatility: A forecasting approach using

the artificial neural network–GARCH model. Expert System with Applications, 42(5),

7245-7251.

Lahmiri, S. (2017). Modeling and predicting historical volatility in exchange rate markets.
re-
Physica A: Statistical Mechanics and its Applications, 471, 387-395.

Li, X. G. & Wu, X. H. (2015). Constructing long short-term memory based deep recurrent

neural networks for large vocabulary speech recognition. 2015 IEEE International

Conference on Acoustics, Speech and Signal Processing (ICASSP): 4520-4524.

lP

Liu, G., & Guo, J. (2019). Bidirectional LSTM with attention mechanism and convolutional

layer for text classification. Neurocomputing, 337, 325-338.

Lu, X., Que, D., & Cao, G. (2016). Volatility forecast based on the hybrid artificial neural
rna

network and GARCH-type models. Procedia Computer Science, 91, 1044-1049.

Maknickienė, N., & Maknickas, A. (2012). Application of neural network for forecasting of

exchange rates and forex trading. Proceedings of the 7th international scientific

conference on business and management, 10–11.

Jou

Nelson, D. B. (1991). Conditional heteroscedasticity in asset returns: A new approach.

Econometrica, 59(2), 347–370.

Nelson, D. M. Q., Pereira, A. C. M., & Oliveira R. A. d. (2017). Stock market's price movement

prediction with LSTM neural networks. International joint conference on neural

networks, 1419-1426.

Parisi, A., Parisi, F., & Díaz, D. (2008). Forecasting gold price changes: Rolling and recursive

neural network models. Journal of Multinational financial management, 18(5), 477–487.

Petropoulos, A., Chatzis, S. P., Siakoulis, V., & Vlachogiannakis, N. (2017). A stacked
27
 Journal Pre-proof

generalization system for automated FOREX portfolio trading. Expert Systems with

Applications, 90, 290-302.

Sardelich, M. & Manandhar, S. (2018). Multimodal deep learning for short-term stock volatility

prediction. arXiv:1812.10479.

of
Schuster, M., & Paliwal, K. K. (1997). Bidirectional recurrent neural networks. IEEE

Transactions on Signal Processing, 45, 2673-2681.

Sermpinis, G., Dunis, C. L., Laws, J., & Stasinakis, C. (2012). Forecasting and trading the

pro
EUR/USD exchange rate with stochastic Neural Network combination and time-varying

leverage. Decision Support Systems, 54(1), 316-329.

Sharda, R., & Patil, R. B. (1992). A connectionist approach to time series prediction: an

empirical test. Journal of Intelligent Manufacturing, 3, 317–323.

re-
Tse, Y. K. (1998). The conditional heterscedasticity of the Yen-Dollar exchange rate. Journal

of Applied Econometrics, 13(1), 49–55.

Wu, Y., X., Wu, Q., -B., & Zhu, J., Q. (2019). Improved EEMD-based crude oil price

forecasting using LSTM networks. Physica A: Statistical Mechanics and its Applications,
lP

516, 114-124.
ˇ ˙
Yazdani-Chamzini, A., Yakhchali, S. H., Volungevic iene , D., & Zavadskas, E. K. (2012).

Forecasting gold price changes by using adaptive network fuzzy inference system. Journal
rna

of Business Economics and Management, 13(5), 994–1010.

Zhang, J., Zhang, Y., & Zhang, L. (2015). A novel hybrid method for crude oil price forecasting.

Energy Economics, 649-659.

Jou

28
 Journal Pre-proof

Highlights:

of
1. We develop a novel hybrid deep learning method to improve forecasts of copper price volatility

2. The hybrid method combines GARCH with neural network models (ANN and LSTM)

pro
3. The empirical results confirm the effectiveness of the proposed method for volatility forecasts

4. The choices between various network configurations are examined to optimize the forecasts

re-
lP
rna
Jou
 Journal Pre-proof

Credit author statement:

Yan Hu: Data curation, Software, Formal analysis, Visualization, Writing - Original draft

of
preparation. Jian Ni: Conceptualization, Methodology, Supervision, Resources, Writing -
Reviewing and Editing. Liu Wen: Software, Validation.

pro
re-
lP
rna
Jou
 Journal Pre-proof

Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.

of
☐The authors declare the following financial interests/personal relationships which may be considered
as potential competing interests:

pro
re-
lP
rna
Jou