Professional Documents
Culture Documents
Abstract— The gold market plays a vital role in the world central bank reserves as it is used as a price controlling lever
economy. Due to its complex and nonstationary nature, pre- and a strategic economic resource [2]. For many investors,
dicting the price of gold is particularly challenging. In this gold is a valuable asset that serves as an investment safe
study, a new hybrid forecasting approach named variational
mode decomposition (VMD)-iterated cumulative sums of squares heaven which can be used to hedge against price inflation and
(ICSS)-bidirectional gated recurrent unit (BiGRU) is proposed other financial risks [3]. In the precious metals market, gold
by integrating BiGRU deep learning model, VMD, and iterated is the most commonly investment commodity [4]. Therefore,
cumulative sum of squares algorithm. The forecasting framework predicting the price of gold accurately is of great importance
is able to extract the inner factors and patterns within the for investors, portfolio managers, as well as policy makers. For
gold futures market movements, decompose its correlation with
external markets and detect shifts within market conditions example, it can create great investment opportunities to benefit
in order to accurately predict price movements in the gold investors as small improvements in forecasting accuracy can
futures market. The experimental results show that the hybrid generate significant profits [5]. However, the gold market is
forecasting approach can improve the prediction performance not only nonstationary and volatile in nature, and it is also
significantly in comparison to the benchmarks. Furthermore, impacted by various factors, such as investor expectations,
we extend the proposed hybrid forecasting approach to generate
trading strategies and test trading performance of the gold related market movements, and political events [6]. Therefore,
futures market. The testing results over an out-of-sample period due to its complexity, predicting trends in the gold market
of 11 years (2008–2019) indicate that the strategy generated based has remained one of the most challenging fields in time-series
on the prediction of the proposed approach displays high levels forecasting [7].
of consistency in generating positive returns and outperforms This article is related to two strands of literature. The
several other common trading strategies under various market
conditions. The approach also shows consistent better results first involves the studies on the influencing factors of the
when generalized to the spot gold market, providing practical gold market. Numerous studies analyzes the effects of input
guidance for minimizing investment risk and hedging strategies variables such as foreign exchange rates and crude oil prices
in the gold commodity market. on the precious metals market [8]. For example, [9], [10] and
Index Terms— Algorithmic trading, bidirectional gated recur- [11] find that exchange rates such as the US Dollar (USD)
rent unit (BiGRU), gold futures price forecasting, variational have significant predictive power in forecasting commodity
mode decomposition (VMD). prices. In addition, [12], [13] and [14] conclude that the price
of oil is a major macroeconomic factor that has significant
I. I NTRODUCTION explanatory power on the price fluctuations of gold, showing
a high correlation of close to 85% between the two markets.
G OLD is one of the most important commodities in the
financial markets, and it plays a vital role in the world
economy [1]. For many countries, gold is a major part of the
Although previous studies have examined the impacts of
these external variables and incorporated them in forecasting
gold price movements, the factors incorporated in these studies
Manuscript received August 12, 2020; revised March 29, 2021; accepted mainly consists of the original observable price variables from
May 16, 2021. Date of publication June 15, 2021; date of current version
December 1, 2021. This work was supported by the National Natural the external markets. However, since the effect duration of
Science Foundation of China under Grant 71801213 and Grant 71988101. different factors on the gold price varies [15], these observable
(Corresponding author: Yunjie Wei.) variables may be insufficient to fully reflect the impact of
Yuze Li is with the Academy of Mathematics and Systems Science, Chinese
Academy of Sciences, Beijing 100190, China, and also with the School of external markets on the gold market. In this article, we select
Economics and Management, University of Chinese Academy of Sciences, four external markets as input variables, including the Brent
Beijing 100190, China (e-mail: richardyz.li@mails.ucas.ac.cn). Crude (LCO), the Western Texas Intermediate Crude (WTI),
Shouyang Wang is with the Academy of Mathematics and Systems Science
and the Center for Forecasting Science, Chinese Academy of Sciences, Beijing the USD Index (DXY) and Gold Spot USD (XAUUSD).
100190, China, and also with the School of Economics and Management, To further extract the hidden relationship between the gold spot
University of Chinese Academy of Sciences, Beijing 100190, China (e-mail: market and the external markets, this article utilizes variational
sywang@amss.ac.cn).
Yunjie Wei is with the Academy of Mathematics and Systems Science mode decomposition (VMD) to decompose the price informa-
and the Center for Forecasting Science, Chinese Academy of Sciences, tion in the selected external markets into subseries of different
Beijing 100190, China (e-mail: weiyunjie@amss.ac.cn). frequencies. The overall correlation between the markets is
Qing Zhu is with the International Business School, Shaanxi Normal
University, Xi’an 710000, China (e-mail: zhuqing@snnu.edu.cn). modeled through multiple correlations between the subseries
Digital Object Identifier 10.1109/TCSS.2021.3084847 of markets. This novel “correlation decomposition” represen-
2329-924X © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
1358 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 6, DECEMBER 2021
tation disentangles the relationship between the gold futures To take advantage of the above mentioned methods and
market and other external markets, which enables the model to overcome their respective limitations, this article proposes
to uncover the hidden relationship between the inner factors a hybrid deep learning neural networks approach by inte-
of gold futures market and the selected external markets. grating signal decomposition technique, structural breakpoint
The second strand of literature involves the forecasting mod- detection and recurrent neural network to predict the gold
els that have been adopted in the financial market predictions. futures price. First, variational mode decompositon (VMD) is
In earlier studies, traditional econometric approaches were used as the signal decomposition technique to decompose the
applied to forecast the commodity prices. Some of the models original complex and nonstationary time series into simple
commonly used include the moving average model (MA), and stationary subseries. The subseries are inputted into the
autoregressive model (AR), autoregressive moving average prediction model as part of the input features, which allows the
model (ARMA) [16], [17], and generalized autoregressive model to capture the inner patterns of the data. Second, several
conditional heteroscedastic model (GARCH) [18], [19]. technical indicators such as the simple moving average (SMA)
The previously mentioned mathematical models have a are included as model inputs. Third, iterated cumulative sums
major drawback that is the models all have the linearity of squares (ICSS) algorithm is utilized to identify the structural
and stationarity assumptions, which is hardly satisfied by the breakpoints in the market. The detected structural breaks
complex and nonstationary nature of the financial market. are incorporated into the prediction model through a binary
Therefore, these assumptions tend to yield poor forecasting dummy variable, which depicts the changes in market condi-
results [20]. To overcome these limitations and take advantage tions. Furthermore, the decomposed subseries of the selected
of the rapid advances in computing technologies, researchers external markets, as well as the respective correlations with
have adopted deep learning approaches in financial market the subseries of the gold futures market, are used as the input
forecasting. Deep learning is a specific type of artificial neural feature to represent the external market factors. Finally, all the
networks (ANNs) consisting of multiple hidden layers. In com- aforementioned features are inputted into the model to forecast
parison to traditional econometric approaches, it contains the prices of the gold futures market. As proposed in [34], time
enhanced functional representation ability and displays supe- series data are bidirectional, which means the current state
rior performance [21], [22]. For example, [23] utilize Deep is the reflection of the past state and the basis of the future
Belief Network (DBN) in combination with the Restricted state. Unlike previous deep learning models that assume the
Boltzmann Machines (RBM) to achieve decent results in stock time series data only contain an unidirectional relationship,
index forecasting. Zhao et al. [24] propose a bootstrapping the approach proposed in this article adopts a bidirectional
aggregation and stacked denoising autoencoders (SDAE) to gated recurrent unit (BiGRU) approach in which information
forecast oil prices. Doering et al. [25] use convolutional is extracted from the forward direction and backward direction.
neural network (CNN) in predicting stock prices based on The proposed approach is completely data-driven and requires
high-frequency data. Deng et al. [26] combine deep direct fewer assumptions, which make it more consistent with reality.
reinforcement (DDR) method, fuzzy deep direct reinforce- To the best of our knowledge, this is the first approach that
ment (FDDR) method together with recurrent deep neural combines signal decomposition, structural break detection, and
network (RDNN) to forecast the Chinese futures market. Fis- bidirectional deep learning structures in the application of
cher and Krauss [27] utilize long short-term memory (LSTM) gold market forecasting. In order to verify the practicality
networks to predict the market trend of S&P 500 stocks of the model, the prediction outputs are used to identify
between 1992 and 2015. In addition, [28] apply deep CNNs the buy, sell, and hold points in the gold futures prices and
to identify buy, sell and hold points for stocks and exchange generate trading strategies. The evaluation results indicate that
traded funds. the proposed approach consistently outperforms some other
To further improve the forecasting ability, an ensemble selected common trading strategies. Furthermore, the approach
learning approach based on the concept of divide and con- is applied to the spot gold market, which displays consistent
quer have been proposed. This approach first decomposes superior trading performances.
the original financial time series into different cycle factors The rest of this article is organized as follows. Section II
then integrate the forecast results of those factors [29], [30]. outlines the theoretical concept design of the model.
Yu et al. [31] adopt a hybrid approach of complementary Section III presents the empirical results and financial
ensemble empirical mode decomposition (CEEMD) and performances. Section IV concludes and provides plans for
extended extreme learning machine (EELM) to forecast the future works.
crude oil price. [15] use CEEMD to decompose the historical
price of international gold into price components and com- II. T HEORETICAL C ONCEPT D ESIGN
bined support vector machine (SVM) and ANN to forecast
the gold prices. Wang et al. [32] propose a novel two-phase This section presents a theoretical concept design for the
decomposition technique by integrating CEEMD and VMD, proposed hybrid VMD-ICSS-BiGRU deep learning approach
and back-propagation neural network (BPNN) to forecast the for forecasting gold futures prices and conducting algorithmic
electricity prices and air quality index in Beijing and Shanghai. trading in the gold futures market. Section II-A provides an
However, one of the drawbacks of this approach is that the overview of the proposed approach, and Sections II-B–II-G
prediction error from decomposition may accumulate, which describe the main steps of the approach, together with the
negatively affects the forecasting ability of the models [33]. related techniques, respectively.
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW HYBRID VMD-ICSS-BiGRU APPROACH FOR GOLD FUTURES PRICE FORECASTING 1359
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
1360 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 6, DECEMBER 2021
where α denotes the quadratic penalty function and λ repre- investment risk. In this article, the standard deviation is used
sents the Lagrangian multipliers. as the measure of volatility. It is calculated as follows in (7):
The solution of (2) is obtained using the alternative direction
n
t=1 (rt − r¯t )
2
method of multipliers (ADMMs) [39]. The estimated subsignal
vt = (7)
modes are shifted from the time domain to the frequency n
domain. The subsignal modes u k and central frequency wk
where rt denotes the log-return of time t, r¯t is the average
are calculated in (3) and (4) as follows:
log-return over the calculated period, n is the number of days
fˆ(w) − û i (w) + λ̂(w) used in the calculation, and v t represents the n day volatility
i =k 2
k (w) =
û n+1 (3) of the asset.
1 + 2α(w − wk )2 In financial analytics, both fundamental analysis and tech-
∞ 2
0 ω|û k (w)| dw nical analysis are widely utilized for developing financial
wkn+1 = ∞ 2
. (4)
time-series models [41]. On the one hand, fundamental analy-
0 |û k (w)| dw
sis evaluates market movements based on influential factors
such as economic conditions, company earnings, liabilities,
C. Structural Break Detection assets, and management. On the other hand, technical analysis
utilizes the past performance of financial prices to make
In order to identify the structural breaks in the time series inferences about future movements. By using technical indi-
that represents significant changes in market conditions, the
cators, it removes the subjectivity in economic factors. In this
ICSS algorithm is employed. Proposed in [40], the ICSS study, several technical indicators are selected to describe the
algorithm is a common approach used for detecting structural
movement trend in the gold futures market, which includes the
breaks in time series. The algorithm assumes that a time series
SMA, exponential moving average (EMA), relative strength
has a stationary variance until whenever a breakpoint occurs. index (RSI), and Williams % R. For each indicator, time
It will then have another stationary variance until the next
intervals from 3 to 20 days are used.
breakpoint. Through the iterative process, all the breakpoints
The SMA is the arithmetic MA of the asset price over a
in the time series can be identified. given period of time. There are many trading strategies widely
Let Ck = kt=1 t2 , k = 1, 2, . . . , T be the cumulated sum
used involving SMAs. Typically, multiple SMAs with different
of square residuals, where {t } denotes the residual series with time intervals can be used together to determine the trend
mean of 0 and variance of σt2 . The statistic Dk can then be
directions of the market. For example, if the asset price is
defined as follows in (5):
above the SMA line, it triggers a “Buy” signal that indicates
Ck k the market is displaying an upward trend. On the contrary,
Dk = − , k = 1, 2, . . . , T, D0 = DT = 0. (5) if the asset price is below the SMA line, it triggers a “Sell”
CT T
signal that indicates the market is showing a downward trend.
If there are no structural breaks in the sample period, there The calculation for SMA is shown below in (8)
are no changes in variance and Dk will remain around 0.
However, if structure breaks occur, Dk will shift away from 0.
a+n
pt
SMA = (8)
Specifically, if the shift exceeds the critical value, a breakpoint n
t=a+1
is identified. In this article, the critical value is set to the value
at the 5% significance level defined in [40]. where pt denotes the asset price at time t and n represents the
number of periods in the calculation.
D. Feature Selection Different from the SMA, the EMA is a weighted MA that
assigns more weighting to recent price data. In comparison to
The features used as the model input include two sets of the SMA, the EMA is more responsive to recent price changes.
components. The first set contains four categories of features It is calculated as shown in (9)
derived from the gold futures market - original price informa-
2
tion, technical indicators, the decomposed closing price sub- EMAt = ( pt − EMAt−1 ) ∗ + EMAt−1 (9)
series and market breakpoints. The original price information, n+1
which consists of the daily log-returns and the 5-day return where EMAt−1 denotes the EMA of the previous period,
volatility, directly reflects the market price movements. The (2/n + 1) is the weighting multiplier applied to the latest price
log-return is the logarithmic rate of return for a particular data.
commodity. It is calculated as follows in (6): The RSI is an oscillator indicator that shows the momentum
of the asset. The RSI oscillates between 0 and 100, which
pt
rt = ln (6) indicates the “overbought” and “oversold” conditions of the
pt−1 asset. Typically, the asset is considered “overbought” if the
where pt denotes the closing price at time t and pt−1 denotes RSI value is above 70 and “oversold” if the RSI value is below
the closing price at time t − 1. 30. The equation for RSI is as follows in (10):
In addition, the volatility of returns measures the disper-
100
sion of the average return of an commodity over a given RSI = 100 − . (10)
period. Typically, the greater the volatility, the higher the 1 + average Gain
average Loss
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW HYBRID VMD-ICSS-BiGRU APPROACH FOR GOLD FUTURES PRICE FORECASTING 1361
The Williams % R is another momentum oscillating indi- represent the external market influences. The two major crude
cator for the asset prices. The value of the indicator oscillates oil indices in the world, The LCO and the WTI Crude (WTI),
between −100 and 0. Similar to RSI, if the value is above −20, are used to represent the oil market. The US DXY, which
the asset is considered “overbought”. If the value is below is a measure of the US dollar value relative to other foreign
−80, the asset is considered “oversold”. The calculation for currencies including the Euro, Japanese Yen, British Pound,
Williams % R is shown below Canadian Dollar, Swedish Krona, and Swiss Franc, is used
Max(high) − pt as a comprehensive indicator for the USD foreign exchange
Williams%R = . (11) market. Finally, the XAUUSD is used to reflect the fluctuations
Max(high) − Min(low)
in the spot gold market.
The VMD is used as the signal decomposition technique The variables incorporated in previous studies mainly con-
to decompose the original complex and nonstationary time sists of the observable market information from the external
series into K simple and stationary subseries. This allows markets. However, this representation may be inadequate to
the model to capture the inner patterns of the data. As stated fully reflect their hidden correlations with the gold futures
by [15], since the effect duration of different factors on the market. Similar to gold futures, the price fluctuations in the
gold price varies, the original gold price time series can external markets also have different time patterns. Therefore,
be decomposed into components of different frequencies to in this study, the external market prices are decomposed into
reflect the different price movement patterns in the market. different subseries, respectively, to extract the inner patterns.
These subseries can effectively capture the effects of different Since the fluctuations of gold prices are affected by factors
factors on the movement of the gold futures market. Thus, of different frequencies, the correlation between different
decomposing the original gold price time series into different subseries of gold futures market and external markets may
subseries can yield meaningful information that allows the vary significantly. For example, the higher frequency gold
model to learn the movement patterns of the market in different futures subseries and higher frequency WTI crude subseries
time intervals. may have different correlations from higher frequency gold
In addition, the ICSS algorithm proposed by [40] is utilized futures subseries and lower frequency WTI oil subseries.
to identify the structural breakpoints in the gold futures price Therefore, in comparison to previous studies a novel correla-
log-returns. The breakpoints represent changes in the market tion representation is proposed in which the overall correlation
return volatility, which is used as a signal to represent shifts between the markets is further decomposed into multiple
in market conditions. Therefore, a binary dummy variable, correlations between the subseries of markets. This correlation
market condition, is incorporated to represent the state of decomposition enables the model to further extract the hidden
the market. The breakpoints are continuously reidentified after relationship between the gold futures market and the selected
each sample. If a breakpoint does not occur, the value of the external markets. In addition, due to the complex nature of
binary variable remains unchanged. If a breakpoint is detected, the market, the correlations between different market subseries
the value of the binary variable is changed, indicating a shift do not remain constant over time. Therefore, a 5-day sliding
in market conditions. The iterative process is repeated until Pearson correlation coefficient is selected as the dynamic rela-
all samples are used. Overall, N structural breakpoints are tionship measure. In this study, the sliding Pearson correlation
detected, which translates to N + 1 market condition values. coefficient is calculated between each decomposed subseries
The first set of features are derived from the gold futures of gold futures market and that of external markets. Each
market. However, only incorporating its own market infor- decomposed external markets subseries is multiplied by its
mation may be insufficient to fully reflect its price move- corresponding correlation coefficient to generate the input
ments. As shown by [42], gold futures prices are found features for the external market component.
to be correlated with that of other major financial markets.
Therefore, movements in the correlated markets may have E. Price Prediction
significant impacts on the fluctuations in gold futures prices. Unlike traditional neural networks that are only able to
Thus, information from several external financial markets use mono-directional time sequential historical information
are incorporated to reflect the hidden relationships. Previous in the iterative direction, a bidirectional neural networks is
studies have investigated the relationship between gold and introduced into the price prediction step. As concluded by [34],
other external variables. Batten et al. [12] conclude that the time series data contains a two-way sequential relationship as
price of oil is a major macroeconomic factor that influences the current state is not only the reflection of historical informa-
the gold price fluctuations. In addition, [13] also discover that tion but also the basis of the future state. Thereby, bidirectional
the price of oil has significant explanatory power on gold price neural networks are more consistent with reality. In this article,
fluctuation behaviors, showing a high correlation between the a BiGRU approach is adopted to extracted information from
two markets. In addition to the price of oil, studies such as the forward direction and backward direction.
[11] show a negative causal relationship between gold prices The bidirectional recurrent neural network (BiRNN) is
and the US dollar. Further, spot gold is often used as an asset proposed by [44]. Different from traditional RNNs that utilize
to hedge against price fluctuations in gold futures. As shown only mono-directional historical information, BiRNN utilizes
by [43], there exist a significant relationship between the price historical information from both directions so that it obtains
volatility of spot gold and gold futures. Based on the above information from past and future states. In addition, in this
evidence, four major financial market indices are selected to article, the traditional RNN cells are replaced by gated current
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
1362 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 6, DECEMBER 2021
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW HYBRID VMD-ICSS-BiGRU APPROACH FOR GOLD FUTURES PRICE FORECASTING 1363
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
1364 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 6, DECEMBER 2021
Fig. 6. Subplot to the left plots the decomposed subseries of the gold futures daily closing prices. The subplot to the right plots the detected structural break
points for one of the data points in the sample.
closing price will increase (decrease) the next day and the data, which means they can effectively capture the sensitive
actual price increases (decreases), then the model prediction is short-term noises in the market.
successful. On the other hand, if the model predicts the asset The right subplot of Fig. 6 illustrates the structural break-
closing price will increase (decrease) the next day and the points detected by the model for one of the data points in
actual price decreases (increases), then the model prediction the sample. The breakpoints, marked by red lines, represent
fails. The greater the DA, the higher the predictive ability of changes in the market return volatility and indicate shifts
the model. The DA is calculated as shown in (23) and (24) in market conditions. Therefore, a dummy variable, market
condition, is used to represent the state of the market. The
1, if (X t − X t−1 ) · ( X̂ t − X t−1 ) ≥ 0 breakpoints are reidentified continuously following each train-
Zt = (23)
0, if (X t − X t−1 ) · ( X̂ t − X t−1 ) < 0 ing day. If a breakpoint does not occur, the value of the market
N condition variable remain unchanged. If a break point occurs,
Zt
DA = i=1
(24) a different market condition value is assigned, indicating a
N shift in market states.
where X̄ t is the mean of the actual value, N is the total number After decomposing the original complex, nonstationary
of data points, X̂ t and X t denote the predicted value and the signals and identifying the shifts in market conditions, the
actual value at time t, respectively. selected input variables are fed into the stacked BiGRU model
to forecast the one-day, two-day, and three-day ahead closing
price for gold futures. Since the closing price of gold futures
D. Empirical Results has close relationships with its prices in the near-past, a 5-day
1) Prediction Performance: In order to avoid the use of successive training window is used. For example, in the case of
future information in the forecasting process and obtain more one-day ahead forecasting, the training data of previous 5 days
realistic performance, we only use VMD to decompose the x t−4 , x t−3 , x t−2 , x t−1 , x t , where x t includes the original inputs,
training dataset. It is important to note that since we adopt technical indicators, the decomposed subseries inner factors,
a rolling testing approach, we perform VMD decomposition the identified market conditions and external market informa-
each time after we incorporate the data from the previous tion of day t, are inputted into the BiGRU model to forecast
testing interval. For illustration purposes, we display the over- the target variable yt+1 , which is the closing price at time t +1.
all decomposition result for the entire data in left subplot of To evaluate the prediction performance of the model, we con-
Fig. 6. As suggested by [47], we decompose the closing price duct both vertical and horizontal performance comparisons
of the asset into 11 subsignal modes of various frequencies. with several benchmark models. For vertical comparisons, the
As shown in Fig. 6, the modes range from low frequency single BiGRU model and the hybrid VMD-BiGRU model are
to high frequency, with M1 being the lowest frequency and selected. Specifically, the BiGRU model excludes the decom-
M11 being the highest frequency. Each subseries represents posed intrinsic modes as its input. The VMD-BiGRU model
a hidden oscillatory factor in the time series. The lowest includes all the components such as the technical indicators,
frequency mode, M1, is able to capture the long-term relative original gold prices, and external market information except
long-term trend of the data. The higher frequency modes, for the identified market breakpoints as its input. For horizontal
M2-M11, depict the relative short-term price fluctuations in the comparison, we construct several benchmark models using
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW HYBRID VMD-ICSS-BiGRU APPROACH FOR GOLD FUTURES PRICE FORECASTING 1365
TABLE I
V ERTICAL F ORECASTING P ERFORMANCE OF D IFFERENT M ODELS
TABLE II
V ERTICAL F ORECASTING P ERFORMANCE BY I NTERVAL
state-of-the-art techniques such as support vector regression the DA of the model increased to 0.7958, which indicates
(SVR), linear regression (LR), ANNs, as well as LSTM. that the model is capable of predicting the marketing trend.
To ensure consistencies between models, we utilize both the Consequently, this suggests that introducing signal decompo-
decomposed time-series and the identified market breakpoints sition into the BiGRU model can significantly improve the pre-
as inputs for all these models. The vertical performance dictive power of the model. Through VMD decomposition, the
comparison results are shown in Tables I and II. The horizontal extracted inner factors as well as the decomposed correlations
performance results are shown in Table III. It is important to between subseries of different markets contain information that
note that all of the normalized prediction results are evaluated is closely related to the fluctuation of the gold futures market.
without the process of unnormalization. Furthermore, when ICSS is introduced to identify changes
In terms of 1-Step ahead forecasting, the RMSE, MAPE, in market conditions, the proposed approach yielded supe-
and MAE of the single BiGRU model are 0.0124, 0.0158, rior results. In comparison to the decomposition-only
and 0.0113, respectively. However, the DA value of the model VMD-BiGRU model, the proposed approach is able to fur-
overall and across all intervals (See Table I) are all below 60%, ther reduce the forecasting errors across all intervals, which
which means that the model is incapable of predicting the indicates a stronger fitting performance. In addition, the DA
market trends. In comparison, when VMD decomposition is further improved to 0.8351, which shows that the proposed
introduced, the hybrid VMD-BiGRU is able to reduce the fore- approach is not only acceptable in mathematical calculations
casting errors and improve the fitting performance. In addition, but also practical in real-world applications.
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
1366 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 6, DECEMBER 2021
TABLE III
H ORIZONTAL F ORECASTING P ERFORMANCE OF D IFFERENT M ODELS
TABLE IV
A NNUALIZED R ETURNS OF D IFFERENT S TRATEGIES
Examining the results for 2-Step ahead and 3-Step proposed VMD-ICSS-BiGRU approach displays an advantage
ahead forecasting, the proposed approach is able to consis- over other strategies by obtaining the highest returns in each
tently reduce the forecasting error and improve the fitting interval. One thing worth noting is that while all other strate-
performance effectively. Therefore, incorporating the signal gies suffered losses in one or multiple intervals, the strategy
decomposition technique and structural breakpoint detection to generated by the proposed approach is able to obtain stable
identify changes in market conditions can effectively capture positive annualized returns across all the testing intervals.
the movement trends in the gold futures market. In addition, As stated by [28], it is particularly challenging to consistently
the results suggest that the bidirectional neural networks struc- beat the market over a long period of time. In this study,
ture is capable of learning the two-way sequential relationship the testing set spans across 11 years from 2008 to 2019.
that exist in the time-series data. During this time, the market has exhibited different fluctuation
As evidence for further verification, as shown in Table III, patterns. Unlike other trading strategies, when the gold futures
when our proposed approach is compared against benchmark market condition is pessimistic or highly volatile, the proposed
models constructed with state-of-art techniques, it is also able model is able to avoid losses and consistently generate positive
to effectively reduce the forecasting errors and improve the profits from the market. During several periods in which other
DA of the forecast. Therefore all of these results suggest that trading strategies performed poorly with negative returns, the
our proposed approach is able to effectively capture the market proposed approach displayed superior performances by gen-
patterns and improve the forecasting performance. erating significant positive profits. Therefore, it is particularly
2) Financial Performance: In order to evaluate the finan- effective and stable when the market condition is pessimistic or
cial performance of the proposed strategy, the hybrid highly volatile. Furthermore, the proposed approach is able to
VMD-BiGRU model, Buy & Hold, 20-day SMA, and 14-day consistently generate positive returns across all testing periods
RSI are used as benchmark to compare the annualized returns regardless of the market condition. Overall, the proposed
across different intervals. Since the single BiGRU model model displayed high levels of stability and in particular,
is incapable of predicting market trends, it is not used to demonstrated strong ability to reduce the negative effects
conduct financial trading. The financial performance of all of suboptimal market conditions. Thus, the model is able
the strategies across different intervals is shown in Table IV, to effectively lower the investment risk in the gold futures,
where the strategy with the highest annualized returns in each thus providing practical significance for investors and financial
interval is labeled in bold. Overall, the annualized return of institutions.
the proposed VMD-ICSS-BiGRU approach across the whole 3) Financial Performance on Spot Gold: To further verify
testing set is 20.41%. In comparison, the VMD-BiGRU model, the robustness of the proposed VMD-ICSS-BiGRU approach,
Buy & Hold, and RSI obtained returns of 12.48%, 2.95%, we apply it to the data from the spot gold market. Similar to the
and 3.73%, respectively. Meanwhile, the SMA experienced gold futures data, the spot gold sample consists of 4981 data
a loss of 7.82%. Examining the results, it is clear that the points from January 10, 2000, to January 25, 2019. Likewise,
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW HYBRID VMD-ICSS-BiGRU APPROACH FOR GOLD FUTURES PRICE FORECASTING 1367
the features for gold futures include original inputs, its tech- R EFERENCES
nical indicators, decomposed subseries inner factors, market
breakpoints, and external market information. In this case, the [1] G. Aye, R. Gupta, S. Hammoudeh, and W. J. Kim, “Forecasting the
price of gold using dynamic model averaging,” Int. Rev. Financial Anal.,
subseries inner factors are extracted from the daily closing vol. 41, pp. 257–266, Oct. 2015.
price of spot gold. External market information used are WTI [2] C. Ciner, C. Gurdgiev, and B. M. Lucey, “Hedges and safe havens: An
Crude, LCO, and gold futures. The initial capital is set to examination of stocks, bonds, gold, oil and exchange rates,” Int. Rev.
Financial Anal., vol. 29, pp. 202–211, Sep. 2013.
$1000 000. In this simulation, 85% of the data are used as the [3] M. F. Ghazali, H. H. Lean, and Z. Bahari, “Is gold a hedge or a safe
training set and the rest are used as the testing set. Therefore, haven? An empirical evidence of gold and stocks in Malaysia,” Int. J.
the testing set includes 747 instances from April 13, 2016, Bus. Soc., vol. 14, no. 3, pp. 428–443, Jan. 2013.
[4] K. Gangopadhyay, A. Jangir, and R. Sensarma, “Forecasting the price of
to January 25, 2019. The model structure remains unchanged gold: An error correction approach,” IIMB Manage. Rev., vol. 28, no. 1,
from the gold futures simulation pp. 6–12, Mar. 2016.
The financial performance of the proposed [5] D. Liu and Z. Li, “Gold price forecasting and related influence factors
analysis based on random forest,” in Proc. 10th Int. Conf. Manage. Sci.
VMD-ICSS-BiGRU approach is compared with that of Eng. Manage., J. Xu, A. Hajiyev, S. Nickel, and M. Gen, Eds. Singapore:
the other four strategies - VMD-BiGRU, Buy & Hold, Springer, 2017, pp. 711–723.
SMA, and RSI. In the testing interval, the proposed [6] R. Gupta, S. Hammoudeh, W. J. Kim, and B. D. Simo-Kengne,
“Forecasting China’s foreign exchange reserves using dynamic model
approach obtained an annualized return of 17.07%, while the averaging: The roles of macroeconomic fundamentals, financial stress
VMD-BiGRU, Buy & Hold, SMA and RSI obtained returns and economic uncertainty,” North Amer. J. Econ. Finance, vol. 28,
of 12.30%, 1.52%, −4.64%, and 5.84%. In comparison, the pp. 170–189, Apr. 2014.
average annualized strategy of the proposed approach not [7] C. Zhang, Z. Pu, and Q. Zhou, “Sustainable energy consumption in
northeast Asia: A case from China’s fuel oil futures market,” Sustain-
only obtained positive profits over the testing period, and ability, vol. 10, no. 1, pp. 1–14, Jan. 2018.
the profits generated are also significantly higher than those [8] L.-T. Zhao, Y. Wang, S.-Q. Guo, and G.-R. Zeng, “A novel method
obtained by the other strategies. Fig. 7 further illustrates based on numerical fitting for oil price trend forecasting,” Appl. Energy,
vol. 220, pp. 154–163, Jun. 2018.
the performance comparison between our strategy and other [9] Y.-C. Chen, K. S. Rogoff, and B. Rossi, “Can exchange rates forecast
strategies. commodity prices?” SSRN Electron. J., vol. 125, no. 3, pp. 1145–1194,
Aug. 2010.
[10] J. C. Reboredo, “Is gold a safe haven or a hedge for the US dollar?
IV. C ONCLUSION Implications for risk management,” J. Banking Finance, vol. 37, no. 8,
In this study, by integrating BiGRU deep learning model, pp. 2665–2676, Aug. 2013.
VMD, and iterated cumulative sum of squares algorithm, [11] B. Mo, H. Nie, and Y. Jiang, “Dynamic linkages among the gold market,
US dollar and crude oil market,” Phys. A, Stat. Mech. Appl., vol. 491,
a hybrid deep learning approach named VMD-ICSS-BiGRU pp. 984–994, Feb. 2018.
is proposed to predict the price movements in the gold [12] J. A. Batten, C. Ciner, and B. M. Lucey, “The macroeconomic determi-
futures market. First, variational mode decomposition (VMD) nants of volatility in precious metals markets,” Resour. Policy, vol. 35,
no. 2, pp. 65–71, Jun. 2010.
is used as the signal decomposition technique to decompose [13] S. Shafiee and E. Topal, “An overview of global gold market and gold
the original complex and nonstationary time series into simple price forecasting,” Resour. Policy, vol. 35, no. 3, pp. 178–189, Sep. 2010.
and stationary subseries. The subseries are inputted into the [14] N. B. Behmiri and M. Manera, “The role of outliers and oil price shocks
on volatility of metal prices,” Resour. Policy, vol. 46, pp. 139–150,
prediction model as part of the input features, which allows the Dec. 2015.
model to capture the inner patterns of the data. Second, several [15] F. Wen, X. Yang, X. Gong, and K. K. Lai, “Multi-scale volatility feature
technical indicators are included as model inputs. Third, ICSSs analysis and prediction of gold price,” Int. J. Inf. Technol. Decis. Making,
vol. 16, no. 1, pp. 205–223, Jan. 2017.
algorithm is utilized to identify the structural breakpoints in [16] A. Parisi, F. Parisi, and D. Díaz, “Forecasting gold price changes: Rolling
the market. Our empirical results indicate that the proposed and recursive neural network models,” J. Multinat. Financial Manage.,
VMD-ICSS-BiGRU approach displays superior prediction per- vol. 18, no. 5, pp. 477–487, Dec. 2008.
[17] T. Kriechbaumer, A. Angus, D. Parsons, and M. R. Casado,
formance results in comparison to other single and hybrid deep “An improved wavelet–ARIMA approach for forecasting metal prices,”
learning models. It is able to consistently generate positive Resour. Policy, vol. 39, pp. 32–41, Mar. 2014.
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.
1368 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 6, DECEMBER 2021
[18] L. Fang, B. Chen, H. Yu, and Y. Qian, “The importance of global [44] M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural net-
economic policy uncertainty in predicting gold futures market volatility: works,” IEEE Trans. Signal Process., vol. 45, no. 11, pp. 2673–2681,
A GARCH-MIDAS approach,” J. Futures Markets, vol. 38, no. 3, Nov. 1997.
pp. 413–422, Mar. 2018. [45] K. Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio, “On the
[19] E. Tully and B. M. Lucey, “A power GARCH examination of the gold properties of neural machine translation: Encoder–decoder approaches,”
market,” Res. Int. Bus. Finance, vol. 21, no. 2, pp. 316–325, Jun. 2007. in Proc. 8th Workshop Syntax, Semantics Struct. Stat. Transl. (SSST-8),
[20] L. Yu, S. Wang, and K. K. Lai, “Forecasting crude oil price with an 2014, pp. 103–111.
EMD-based neural network ensemble learning paradigm,” Energy Econ., [46] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of
vol. 30, no. 5, pp. 2623–2635, Sep. 2008. gated recurrent neural networks on sequence modeling,” in Proc. NIPS
[21] Y. Bengio, “Learning deep architectures for AI,” Found. Trends Mach. Workshop Deep Learn., Dec. 2014, pp. 1–9.
Learn., vol. 2, no. 1, pp. 1–127, 2009. [47] Q. Zhu, F. Zhang, S. Liu, Y. Wu, and L. Wang, “A hybrid VMD–BiGRU
[22] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, model for rubber futures time series forecasting,” Appl. Soft Comput.,
pp. 436–444, May 2015. vol. 84, Nov. 2019, Art. no. 105739.
[23] T. Kuremoto, S. Kimura, K. Kobayashi, and M. Obayashi, “Time
series forecasting using a deep belief network with restricted Boltzmann
machines,” Neurocomputing, vol. 137, no. 15, pp. 47–56, Aug. 2014.
[24] Y. Zhao, J. Li, and L. Yu, “A deep learning ensemble approach for crude
oil price forecasting,” Energy Econ., vol. 66, pp. 9–16, Aug. 2017. Yuze Li is currently pursuing the master’s degree in
[25] J. Doering, M. Fairbank, and S. Markose, “Convolutional neural net- management science and engineering with the Acad-
works applied to high-frequency market microstructure forecasting,” in emy of Mathematics and Systems Science, Chinese
Proc. 9th Comput. Sci. Electron. Eng. (CEEC), Sep. 2017, pp. 31–36. Academy of Sciences, Beijing, China.
[26] Y. Deng, F. Bao, Y. Kong, Z. Ren, and Q. Dai, “Deep direct rein- His research interests include economic modeling,
forcement learning for financial signal representation and trading,” IEEE analysis, forecasting and game theory. He has pub-
Trans. Neural Netw. Learn. Syst., vol. 28, no. 3, pp. 653–664, Mar. 2017. lished in journals including Nature Communications,
[27] T. Fischer and C. Krauss, “Deep learning with long short-term memory European Journal of Operational Research, and
networks for financial market predictions,” Eur. J. Oper. Res., vol. 270, Energy Economics.
no. 2, pp. 654–669, Oct. 2018.
[28] O. B. Sezer and A. M. Ozbayoglu, “Algorithmic financial trading with
deep convolutional neural networks: Time series to image conversion
approach,” Appl. Soft Comput., vol. 70, pp. 525–538, Sep. 2018.
[29] Q. Wang, W. Xu, X. Huang, and K. Yang, “Enhancing intraday
stock price manipulation detection by leveraging recurrent neural net-
works with ensemble learning,” Neurocomputing, vol. 347, pp. 46–58, Shouyang Wang received the Ph.D. degree in oper-
Jun. 2019. ations research from the Institute of Systems Sci-
[30] J. Xiao, Y. Tian, L. Xie, X. Jiang, and J. Huang, “A hybrid classification ence, Chinese Academy of Sciences, Beijing, China,
framework based on clustering,” IEEE Trans. Ind. Informat., vol. 16, in 1986.
no. 4, pp. 2177–2188, Apr. 2020. He is currently a Bairen Distinguished Professor
[31] L. Yu, Z. Wang, and L. Tang, “A decomposition–ensemble model with of management science with the Academy of Math-
data-characteristic-driven reconstruction for crude oil price forecasting,” ematics and Systems Science, Chinese Academy of
Appl. Energy, vol. 156, pp. 251–267, Oct. 2015. Sciences. He has received many research related
[32] D. Wang, S. Wei, H. Luo, C. Yue, and O. Grunder, “A novel hybrid awards and honors. He has published 35 mono-
model for air quality index forecasting based on two-phase decom- graphs and published more than 330 articles in
position technique and modified extreme learning machine,” Sci. Total international academic journals. He is/was a Coed-
Environ., vol. 580, pp. 719–733, Feb. 2017. itor of 16 journals and a Guest Editor of special issues/volumes of more
[33] L. Tang, S. Wang, K. He, and S. Wang, “A novel mode-characteristic- than 15 journals. His research interests include decision analysis, risk man-
based decomposition ensemble model for nuclear energy consumption agement, economic analysis, and forecasting.
forecasting,” Ann. Oper. Res., vol. 234, no. 1, pp. 111–132, Nov. 2015.
[34] R. Kadari, Y. Zhang, W. Zhang, and T. Liu, “CCG supertagging via bidi-
rectional LSTM-CRF neural architecture,” Neurocomputing, vol. 283,
pp. 31–37, Mar. 2018.
[35] K. Dragomiretskiy and D. Zosso, “Variational mode decomposition,” Yunjie Wei received the Ph.D. degree in manage-
IEEE Trans. Signal Process., vol. 62, no. 3, pp. 531–544, Feb. 2014. ment science and engineering from the Academy of
[36] Y. Wang and R. Markert, “Filter bank property of variational Mathematics and Systems Science, Chinese Acad-
mode decomposition and its applications,” Signal Process., vol. 120, emy of Sciences, Beijing, China, in 2017, and the
pp. 509–521, Mar. 2016. Ph.D. degree in management science from the City
[37] W. Liu, S. Cao, and Y. Chen, “Applications of variational mode University of Hong Kong, Hong Kong, in 2018.
decomposition in seismic time-frequency analysis,” Geophysics, vol. 81, She is currently an Assistant Professor with the
no. 5, pp. V365–V378, Sep. 2016. Academy of Mathematics and Systems Science,
[38] C. Zhang, J. Zhou, C. Li, W. Fu, and T. Peng, “A compound structure of Chinese Academy of Sciences. She has published
ELM based on feature selection and parameter optimization using hybrid over 20 articles in international academic journals.
backtracking search algorithm for wind speed forecasting,” Energy Her research interests include economic modeling,
Convers. Manage., vol. 143, pp. 360–376, Jul. 2017. and analysis and forecasting.
[39] M. R. Hestenes, “Multiplier and gradient methods,” J. Optim. Theory
Appl., vol. 4, pp. 303–320, Nov. 1969.
[40] C. Inclán and G. C. Tiao, “Use of cumulative sums of squares for
retrospective detection of changes of variance,” J. Amer. Stat. Assoc.,
vol. 89, no. 427, pp. 913–923, Sep. 1994. Qing Zhu is a Professor of finance with the Interna-
[41] R. C. Cavalcante, R. C. Brasileiro, V. L. F. Souza, J. P. Nobrega, and tional Business School, Shaanxi Normal University,
A. L. I. Oliveira, “Computational intelligence and financial markets: A Xi’an, China, and a Postdoctoral Researcher in data
survey and future directions,” Expert Syst. Appl., vol. 55, pp. 194–211, science with the School of Management of Xi’an
Aug. 2016. Jiaotong University, Xi’an. His research interests are
[42] K. Kanjilal and S. Ghosh, “Dynamics of crude oil and gold price big data and soft computing, machine learning, and
post 2008 global financial crisis—New evidence from threshold vector neural networks.
error-correction model,” Resour. Policy, vol. 52, pp. 358–365, Jun. 2017.
[43] W. Kristjanpoller and M. C. Minutolo, “Gold price volatility: A fore-
casting approach using the artificial neural network–GARCH model,”
Expert Syst. Appl., vol. 42, no. 20, pp. 7245–7251, Nov. 2015.
Authorized licensed use limited to: National Institute of Technology Goa. Downloaded on February 23,2024 at 19:55:05 UTC from IEEE Xplore. Restrictions apply.