Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 00 (2018) 000–000
Procedia
Procedia Computer
Computer Science
Science 00(2018)
139 (2018)25–32
000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

The International Academy of Information Technology and Quantitative Management, the Peter
The International Academy Kiewit
of Information
Institute,Technology
University and Quantitative Management, the Peter
of Nebraska
Kiewit Institute, University of Nebraska
Forecasting
Forecasting Exchange
Exchange Rate
Rate Value
Value at
at Risk
Risk using
using Deep
Deep Belief
Belief
Network Ensemble based Approach
Network Ensemble based Approach
Kaijian Hea,b,∗ b c d e
a,b,∗, Lei Jib , Geoffrey K.F. Tsoc , Bangzhu Zhud , Yingchao Zoue
a Hunan
Kaijian He , Lei Ji , Geoffrey K.F. Tso , Bangzhu Zhu , Yingchao Zou
Engineering Research Center for Industrial Big Data and Intelligent Decision Making, Hunan University of Science and Technology,
a Hunan Engineering Research Center for Industrial Big Data and Intelligent Decision Making, Hunan University of Science and Technology,
Xiangtan 411201, China
Xiangtan
b School of Business, Hunan University 411201,
of Science China
and Technology, Xiangtan, 411201, China
c Department b School of Business, Hunan University of Science and Technology, Xiangtan, 411201, China
of Management Sciences, City University of Hong Kong, Tat Chee Avenue, Kowloon Tong, Hong Kong
c Department of Management
d Business Sciences, City University of Hong Kong,&Tat Chee Avenue, Kowloon Tong, Hong Kong
School, Nanjing University of Information Science Technology, Nanjing 210044,China
e Collegedof
Business School,
Information Nanjing
Science andUniversity of Information
Technology, Scienceof&Chemical
Beijing University Technology, Nanjing Beijing
Technology, 210044,China
100029, China
e College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China

Abstract
Abstract
In this paper, we propose a new Value at Risk estimate based on the Deep Belief Network ensemble model with Empirical
In this
Mode paper, we propose
Decomposition (EMD)atechnique.
new ValueIt at Risk estimate
attempts based
to capture the on the Deepdata
multi-scale Belief Network
features with ensemble
the EMD-DBNmodelensemble
with Empirical
model
Mode Decomposition
and predict (EMD) technique.
the risk movement It attempts
more accurately. to capture
Individual thecomponents
data multi-scale data features with
are extracted usingtheEMD
EMD-DBN ensemble
model while model
individual
and predict the risk movement more accurately. Individual data components are extracted using EMD
forecasts can be calculated at different scales using ARMA-GARCH model. The DBN model is introduced to search for the model while individual
forecasts can be calculated
optimal nonlinear at different
ensemble weights scales using
to combine ARMA-GARCH
the individual forecastsmodel. The scales
at different DBN model
into theisensembled
introducedexchange
to searchrate
forVaR
the
optimal
forecasts.nonlinear
Empiricalensemble weights
studies using to combine
major exchange therates
individual
confirmforecasts
that the atproposed
differentmodel
scalesdemonstrates
into the ensembled exchange
the superior rate VaR
performance
forecasts.
compared Empirical studies models.
to the benchmark using major exchange rates confirm that the proposed model demonstrates the superior performance
compared to the benchmark models.

©c 2018
2018 The
The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
c 2018an The Authors. Published by Elsevier B.V.
This is
This is an open
open access
access article
article under
under the
the CC
CC BY-NC-ND
BY-NC-ND license
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is
Peer an open
review
review access
under
under article under
responsibility
responsibility thethe
ofof
the CCscientific
BY-NC-ND
scientific license
committee
committee (http://creativecommons.org/licenses/by-nc-nd/4.0/)
of The International
of The Academy
International of Information
Academy Technology
of Information and Quan-
Technology and
Peer review
Quantitative
titative under
Management, responsibility
Management, of the
the Peter
the Peter Kiewit scientific
Kiewit committee
Institute,
Institute, ofofNebraska.
University
University The International Academy of Information Technology and Quan-
of Nebraska.
titative Management, the Peter Kiewit Institute, University of Nebraska.
Keywords:
Keywords:
Exchange rate forecasting; Empirical Mode Decomposition; Deep Belief Network; Value at Risk
Exchange
JEL: F31; rate
C45;forecasting;
C53 Empirical Mode Decomposition; Deep Belief Network; Value at Risk
JEL: F31; C45; C53

1. Introduction
1. Introduction
Exchange market is one of the most lucrative and actively traded financial markets. Exchange rate determined
by Exchange
the marketmarket
supplyisand
onedemand
of the most lucrative also
for currencies and actively traded
reflects the financialfactors
influencing markets. Exchange
in the ratemarket.
exchange determined
It is
by
one of the key macroeconomic factors linking the global financial system [1, 2]. Due to the global exposure ofItthe
the market supply and demand for currencies also reflects the influencing factors in the exchange market. is
one of the key macroeconomic factors linking the global financial system [1, 2]. Due to the global exposure of the

∗ Corresponding author. Tel.: +86-731-58290306 ; fax: +86-731-58290306.

∗ Corresponding
E-mail address:author. Tel.: +86-731-58290306 ; fax: +86-731-58290306.
[email protected]
E-mail address: [email protected]
1877-0509  c 2018 The Authors. Published by Elsevier B.V.
1877-0509
1877-0509 ©
c 2018
2018 The
The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
This
This is
is an
anopen
openaccess
accessarticle
articleunder
underthethe
CCCCBY-NC-ND
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an
Peer review open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer review under
underresponsibility
responsibilityofof
thethe
scientific committee
scientific committeeof The International
of The Academy
International of Information
Academy Technology
of Information and Quantitative
Technology Manage-
and Quantitative
Peer review
ment,
Management, under
the Peter the responsibility
Kiewit
PeterInstitute, of the scientific
KiewitUniversity
Institute, of committee
Nebraska.
University of The International Academy of Information Technology and Quantitative Manage-
of Nebraska.
ment, the Peter Kiewit Institute, University of Nebraska.
10.1016/j.procs.2018.10.213
26 Kaijian He et al. / Procedia Computer Science 139 (2018) 25–32
2 K. HE, L. JI, G. TSO, B. ZHU, Y. ZOU / Procedia Computer Science 00 (2018) 000–000

exchange markets to wide range of investors, significant price fluctuations and risks have been observed in the market
trading. However, the interaction among these factors are so complicated that it exceeds the computational boundary
of the investors to conduct exhaustive investigations into the relationship between them and forecast the exchange rate
movement accurately. The perceived market risk can be attributed to a diverse range of influencing factors, mostly
concerning the macroeconomic indicator, etc. [1, 2]. The accurate risk measurement and modeling critically affect the
theoretical research in international economics and finance such as the derivative pricing risk measurement.
In this paper, the downside risk in the exchange market is measured using the widely adopted Value at Risk (VaR)
concept. VaR is defined as the particular quantile value of the exchange rate distributions over the given time horizon
at the specified confidence level. VaR estimation models are usually classified into the parametric and non-parametric
categories in the mainstream literature [3]. The parametric models include Exponential Weighted Moving Average
model, GARCH model while the non-parametric models include historical simulation and Monte Carlo simulation
models [3]. It summarizes the downside risk over a particular investment period into a single statistic number [3]. To
estimate VaR more reliably, researches have been directed toward analyzing the fundamental and technical factors
behind the exchange rate fluctuations, with the aims to exploit the key risk factors for more accurate risk modeling.
Researches on analyzing and exploiting both factors for the risk measurement purposes have received continuous
research attentions in the literature. But since the seminal work by [4], it has been shown that the major econometric
models failed to predict the exchange rate changes better than the naive Random Walk (RW) model. So far there is
no conclusive evidence on which exchange rate model is the best method to model the complex fluctuations in the
exchange rate movement. Empirical studies show that the influence factors in the exchange markets are difficult to
identify and they typically have nonlinear dynamic influences on the exchange rate movement over time. This gives
rise to the necessity and renewed interests to model and control the significant risk in the exchange markets.
Recently multiscale ensemble models have received significant research interests in the literature [5, 6]. Current
approaches have used the wavelet analysis and Empirical Mode Decomposition to decompose the financial data into
its constituent components and use the ensemble models to integrate forecasts in those constituent components [6,
7]. These models develop with the recent advancement of two recent researches development, i.e. EMD model and
Artificial Intelligence (AI) ensemble model. On one hand, EMD was first proposed by Huang in 1998 and quickly
found its applications in geophysics, vibration engineering, biomedicine, and other fields [8, 5]. [9] introduced EMD
as a new tool for analyze the statistical characteristics of nonlinear and non-stationary data in the financial market. On
the other hand, AI based ensemble approaches have served as a complementary approach to exchange rate modeling.
Given the lack of definitive analytical tools for the market dynamics, AI can be used to model the dynamically
changing weights in the ensemble models for risk estimate [10, 11, 12]. Till now limited positive performance have
highlighted its significance. For example, empirical studies in the exchange rate and crude oil forecasting field have
shown the higher level of forecasting accuracy [7].
As far as the AI model is concerned, the recent emergence of Deep Learning (DL) model and its success in
physical disciplines compared to traditional neural network model showcase its potential to enhance the ensemble risk
modeling accuracy in the exchange rate. For example, [13] used the deep belief restricted Boltzmann machine to learn
and model the complex relationship between the real estate unit sales price and large number of economic variables.
They found the improved accuracy and reduced computing time for their proposed approach. [14] used the deep belief
network to model and forecast the hourly electricity load and found the improved forecasting accuracy. [15] used the
deep neural networks to forecasting the S&P 500 stock index movement. The improved forecasting accuracy shows
that there is exploitable forecasting opportunities in the stock market. Till now only limited exploratory attempts
using the DBN model have been conducted in the exchange markets, mostly for the forecasting purpose. Positive
performance improvements have already accumulated, etc. For example, [16] found that DBN model provided the
superior exchange rate forecasting accuracy than the neural network model.
In this paper, we propose a EMD-DBN ensemble Value a Risk model that takes advantage of the DBN ensemble
model and EMD model. It relies on the EMD model to recover the influencing factors for the exchange markets at
different scales. It uses the DBN model to integrate the risk forecasts of a series of extracted constituent factors in
the multi scale domain to produce the optimal risk estimate. We have conducted the empirical analysis applying the
proposed model in the exchange markets. Experiment results show the superior performance of the proposed model
and the validity of the extracted multi scale components. The major contribution of the paper is the proposition of DBN
ensemble risk estimate model with EMD model. It introduces DBN model in the risk modeling to take full advantage
 Kaijian He et al. / Procedia Computer Science 139 (2018) 25–32 27
K. HE, L. JI, G. TSO, B. ZHU, Y. ZOU / Procedia Computer Science 00 (2018) 000–000 3

of its nonlinear adaptive learning ability. As the DBN model is known for its capability to extract the nonlinear data
pattern, it is used as the basis for the nonlinear ensemble model. As the DBN model is known to be less sensitive to
the noise disruption, it contributes to the risk measurement literature to further performance improvement in the multi
scale risk forecasts.
The rest of the paper is organized as follows. Section 2 provides a brief illustration of the DBN model. It further
explains in detail the value at risk estimation model using the DBN ensemble model with EMD technique. Section
3 explains the results of the empirical studies useful using the comprehensive exchange rate dataset to evaluate the
model performance. Section 4 summarizes the findings of the paper.

2. Methodology

2.1. Deep Belief Network (DBN)

Deep Learning (DL) has attracted significant research attentions from both academics and industry since it was
originally proposed in 2006 [17]. Deep Belief Network (DBN) is a generative graphical model in the general deep
learning framework. Like traditional shallow neutral network, the neurons in DBN have activation function and pro-
cess the information. As the number of layers increases rapidly from the shallow neural network to the deep neural
network, the number of the parameters involved in training deep neural network increases exponentially. To reduce
the over fitting problem, the pre-training using a stack of a number of restricted Boltzmann machines(RBMs) is in-
troduced in DBN. RBM is used as the pre-training model to learn the hidden data feature in an unsupervised learning
process. A RBM is a neural network that consists of two layers called visible (input) layer and hidden layer. In RBM,
neurons in the different levels of layers have mutual undirected connection while the neurons in the same level of
layers are independent [18]. In the structure of RBM , v and h represent the state of the neurons of visible layer and
hidden layer respectively. When a number of RBMs are stacked in DBN, the hidden layer of the previous RBM serves
as the visible layer of the following RBM [19].
The typical structure of DBN is illustrated in figure 1 [17, 20, 21].

Fig. 1. The general architecture of DBN

The training process of DBN model can usually be divided into pre-training phase with an unsupervised learning
process and fine tuning phase with a supervised learning process. In the pre-training phase ,the initial weights of
the network are optimized so that the input data can be reconstructed. That way, the important data features in the
input data are extracted in the RBM. The RBMs are sequantially trained with an algorithm called greedy layer-wise
28 Kaijian He et al. / Procedia Computer Science 139 (2018) 25–32
4 K. HE, L. JI, G. TSO, B. ZHU, Y. ZOU / Procedia Computer Science 00 (2018) 000–000

unsupervised learning algorithm. In the fine-tuning phase, some label units are added to the topmost layer from the pre-
training phase. The initial weights from the pre-training phase are trained and adjusted by back-propagation algorithm
in a supervised learning process. The global optimal weights of deep belief network are obtained [22, 23].
In the first pre-training step, the parameters of RBM is updated by the process of unsupervised learning. The
weights of whole RBMs of DBN can be updated sequentially with a greedy layer-wise pre-training process. Given
that vi and h j as the state of visible unit i and hidden unit j, the probability distribution of the unit is calculated by the
entropy function since the RBM is an energy-based model [24]. The energy function is calculated as in Eq. (1) [20].

 N
M  M
 N

E(v, h; θ) = − wRij vi h j − ai vi − b jh j (1)
i=1 j=1 i=1 j=1

Where wRij is the connecting weight that connects the neurons i in the visible layer and the neurons j in the hidden
layer. The ai is the bias of the visible neuron i, the b j is the bias of the hidden neuron j. θ is the parameter set that
θ = {W, a, b}. The M and N are represent the number of neurons in the visible layer and hidden layer [25].
In RBM, the joint probability distribution of RBM with respect to units in the visible layer and hidden layer can be
defined by the energy function as in Eq. (2) [20].

e−E(v,h;θ)
p(v, h; θ) = (2)
Z


Where Z is a partition function that used for normalization Z = v,h e−E(v,h;θ) .
Because the joint probability distribution of RBM can be calculated as in Eq. (2), the marginal probability of visible
neurons in RBM is calculated by summing the probability distribution of the all hidden units, as in Eq. (3) [20].

 
h e−E(v,h;θ)
P(v; θ) = p(v, h; θ) = (3)
h
Z

To update the weight, the derivative of Eq. (4) is taken with respect to the weight. The weights of RBM is updated
as in Eq. (4) [14].

∂ log p(v; θ)
∆wRij = µ( ) = µ(Edata (vi h j ) − Emodel (vi h j )) (4)
∂wRij

Where µ denotes the learning rate, Edata (vi h j ) represents the expected values of the observed data in the training
data. Emodel (vi h j ) represents the expected values of the sampling data in the model. In the literature, Gibbs sampling is
usually used to calculate the approximate value for Emodel (vi h j ) [20]. For example, Contrastive Divergence algorithm
based on the Gibbs sampling has been shown to be an effective approach [22]. Through the above training process,
the weights of the RBM can be update [22].
The whole initial weights of DBN can be updated by sequential train RBMs using the above method. After the
unsupervised pre-training step, the supervised fine-tuning step with back-propagation algorithm is used to adjust the
weights slightly obtained from unsupervised learning process to obtain the optimal weights of DBN.
 Kaijian He et al. / Procedia Computer Science 139 (2018) 25–32 29
K. HE, L. JI, G. TSO, B. ZHU, Y. ZOU / Procedia Computer Science 00 (2018) 000–000 5

2.2. a Deep Belief Network Ensemble Value at Risk Estimate

We construct an ensemble Value at Risk model based on Empirical Mode Decomposition (EMD) and the Deep
Belief Network (DBN) model. It consists of a series of steps to transform and model the data characteristics.
(1) Firstly, we assume that there is a multi scale risk structure in the exchange rate movement. Some of the risk
factors have simultaneous impact across different markets while some of the risk factors have limited impact on single
market. Given exchange rate time series data xi , i = 1, 2, ..., N, EMD model is used to calculate the decomposed
components I MF at different scales up to the maximal scale N and the residual e, as in Eq. (5).

N

xt = I MFi,t + et (5)
i=1

This is one of the key step in the model. It transforms the exchange rate data into a series of data components,
upon which further ensemble modeling is performed. The multi scale transformation is done using the Empirical
mode decomposition (EMD) model. The Empirical Mode Decomposition (EMD) method decomposes the original
signal into a number of mutually independent Intrinsic Mode Functions (IMFs) through the sifting process [8]. Unlike
wavelet analysis, no basis function needs to be set during the decomposition process. Compared with wavelet analysis,
the empirical and adaptive nature of the EMD basis make it more suitable for accurately describing the physical
characteristics of the signal when dealing with non-stationary nonlinear signals. It decomposes the data into several
intrinsic mode functions (IMF) at different scales with unique characteristic. Although the empirical approach of EMD
decomposition can’t guarantee the mutual independence of IMFs mathematically, the correlation between IMFs are
supposed to be very low in the sifting process. When applied to the exchange rate modeling, IMFs at different scales
represent the influence of different investment factors on the market volatility and risk exposure.
(2) Secondly, we estimate the ARMA-GARCH model for individually extracted data components using the model
training data. The conditional mean and conditional standard deviation are forecasted using ARMA-GARCH model,
with normal assumption [26]. The model lags for both ARMA and GARCH models are determined using the model
tuning data set by minimizing the AIC information criteria [26].
(3) Thirdly, we assume that there is nonlinear relationship between the individual risk factors and the total risk
estimate. We use the DBN based nonlinear ensemble model to combine individual forecasts of both conditional mean
and conditional standard deviation at different scales. This would produce the ensemble forecasts of conditional mean
and conditional standard deviation.
(4) Fourthly, with the estimated ensemble conditional mean and conditional standard deviation, we forecast the
future VaR using the standard parametric approach [3]. Assuming the normal distribution, VaR at confidence level cl
is calculated as in Eq. (6) [3].

VaRa = −µ + σZa (6)

Where a = 1 − cl. Za is the normal variate at quantile a.µ and σ is the conditional mean and conditional standard
deviation.

3. Empirical Studies

In this section, we conducted the empirical investigation of the performance of the proposed model against the
benchmark model. ARMA-GARCH model is used as the benchmark model since it is one of the most robust models
that provides the baseline performance in the literature [7]. During the empirical evaluation, we have used the extensive
market dataset across major exchange markets. These exchange rates include Australian Dollar against Dollar (AUD),
30 Kaijian He et al. / Procedia Computer Science 139 (2018) 25–32
6 K. HE, L. JI, G. TSO, B. ZHU, Y. ZOU / Procedia Computer Science 00 (2018) 000–000

Dollar against Canadian Dollar (CAD), Dollar against Swiss Franc (CHF), and Euro against Dollar (EURO). All data
were downloaded from the Quandl, which stores a large number of data from different sources and makes it easy for
investors to access the data publicly [27].
The dataset covers the period from 23, July 2007 to 3, August 2018. It contains 2772 daily observations. When
constructing the dataset, the original data has been transformed using the log difference transformation. In this study,
the first 70% of the dataset was used as the model training set, to estimate the ARMA-GARCH model. The rolling
window was set to the length of the training set. The holding period was assumed to be 1 day. The rolling time window
was moved one step ahead each time. The next 21% of the dataset is used as the model tuning dataset, to train and
estimate the DBN model. The remaining 9% of the dataset, i.e. 250 observations, was reserved as model test dataset,
to evaluate the model performance. The historical data is decomposed into 8 groups of mutually independent IMFs
using the EMD method. The software implementations include tensorflow firm Google, the Python, R and the matlab
software, as well as various packages including Keras, numpy,, etc [28]. The DBN is implemented using the software
packages provided by [29].
Firstly we calculate the descriptive statistics and statistical test of data distribution characteristics. Results are listed
in table 1.

Table 1. Descriptive statistics and statistical tests

Statistics AUD CAD CHF Euro
Mean(x10-5) -7.2658 9.6327 -10.6395 -13.6789
Standard Deviation 0.0099 0.0068 0.0080 0.0067
Skewness -0.6691 -0.0954 -1.7297 0.1900
Kurtosis 15.2254 8.7051 48.4565 6.0503
JB Test 0.001 0.001 0.001 0.001
BDS Test 0 0 0 0.0109

Results in Table 1 show that the exchange rate movement is characterized by non-normal distribution, with complex
nonlinear behavior. The rejection of the null hypothesis for Jarque-Bera test confirms the non-normal distribution
of the exchange rate movement. The rejection of the null hypothesis of BDS tests confirms that the exchange rate
movement are not independently distributed. AUD market has the highest level of standard deviation while Euro
market has the lowest level of standard deviation. Exchange rate has negative skewness in all exchange markets
expect for Euro exchange market. Heavy tails with prevalence of extreme or transient events are observed in exchange
rate distribution in all exchange markets as indicated by signfiicant kurtosis value. Among four exchange markets, the
kurtosis value is especially large in the Swiss Frac market.
Then we conduct the experimenting to forecast the one step ahead VaR using different models. When estimating
the DBN model, the empirical volatility is chosen as the proxy for the latent volatility level. It was calculated over
a one year running time window, i.e. 250 trading days. The VaR model performance is evaluated using the back
testing procedure [3]. The main performance measure is the number of exceedance calculated from the back testing
procedure. We also calculate the p values of the test statics for the unconditional coverage test using the number
of exceedances [3]. For benchmark ARMA-GARCH model, lag one is used and the model form is ARMA(1,1) -
GARCH(1,1). As for DBN network, the network structure is 1 input layer, 3 hidden layers and 1 output layer, with
8-50-50-50-1 structure. For the initial tuning phase, the learning rate for is 0.005 and the maximal number of iterations
is 40. For the fine tuning phase, the learning rate is 0.01 and the maximal number of iterations is 300. The activation
function is Rectified Linear Unit (ReLU). The parameters for DBN model is trained on a interval of 126 trading days,
i.e. the network is trained twice a year.
The forecasting accuracy in terms of different performance measures are listed in table 2.
 Kaijian He et al. / Procedia Computer Science 139 (2018) 25–32 31
K. HE, L. JI, G. TSO, B. ZHU, Y. ZOU / Procedia Computer Science 00 (2018) 000–000 7

Table 2. The model performance using the out-of-sample test data

Models NAUD NCAD NCHF NEuro pAUD pCAD pCHF pEuro
AG95% 13 16 7 8 0.8853 0.3294 0.0828 0.1632
AG97.5% 8 11 5 5 0.4965 0.0817 0.6000 0.6000
AG99% 4 4 3 1 0.3805 0.3805 0.7580 0.2781
AG Average 8.3333 10.3333 5 4.6667 0.5875 0.2639 0.4803 0.3471
DBN95% 15 14 9 10 0.4812 0.6691 0.2860 0.4529
DBN97.5% 7 11 7 6 0.7656 0.0817 0.7656 0.9188
DBN99% 3 6 3 2 0.7580 0.0594 0.7580 0.7419
DBNAverage 8.3333 10.3333 6.3333 6 0.6683 0.2701 0.6032 0.7045

Where AGcl , DBNcl ,cl = {95%, 97.5%, 99%, Average} refers to VaR estimate using ARMA-GARCH model and
Deep Belief Network model at cl confidence level. The Average refers to the average value of VaR estimates at cl
confidence levels.
In terms of the average p value, both the proposed model and ARMA-GARCH model have passed the unconditional
coverage test as no p value is less than 0.05 at all confidence levels. As for ARMA-GARCH model, the risk forecasts
are more aggressive in AUD and CAD markets since the number of exceedances are less than the theoretical values.
They are more conservative in CHF and Euro markets since the number of exceedances are larger than the theoretical
values. Compared with the benchmark model ARMA-GARCH model, risk forecasts produced by DBN based model
are more conservative in AUD and CAD markets at 95% and 97.5% confidence levels and more aggressive in CHF
and Euro markets at all confidence levels. Thus the DBN based model produces generally more optimal forecasts. In
terms of the average p value, the DBN based model performs better than ARMA-GARCH model. We found that using
the DBN model leads to the best out-of-sample forecasting performance. The use of nonlinear ensemble integrating
forecasts at 8 different scales shows that forecasts at different scales do not contribute the same weight in the nonlinear
ensemble algorithm. In the meantime, the DBN model is trained on a less regular basis. The model parameters remain
the same between two different trainings. The resulting improved risk estimation accuracy implies that multiscale
structure for the contribution of forecasts at different scales are stable over a half year period.

4. Conclusions

In this paper, we have proposed a DBN based nonlinear ensemble model with EMD technique, for estimating value
at risk. DBN has been used to aggregate the ensemble multiscale risk forecasts in the foreign exchange market. Positive
performance improvement in risk estimates have been observed. Results in this paper have demonstrated that the use of
DBN model could identify more optimal ensemble weights and better integrate the partial information from extracted
risk estimates. This approach leads to the forecasting accuracy improvement of the multiscale VaR estimate models.
Work in this paper implies that new innovative deep learning model can take into account the domain knowledge in
the risk estimate field to gain better insights in the risk estimate field and achieve the improved forecasting accuracy.

Acknowledgments

This work is supported by the National Natural Science Foundation of China (NSFC No. 71671013), and Human-
ities and Social Sciences Youth foundation of Ministry of Education of China (No. 16YJC790026).

References

[1] A. Abhyankar, L. Sarno, G. Valente, Exchange rates and fundamentals: Evidence on the economic value of predictability, Journal of Interna-
tional Economics 66 (2) (2005) 325–348.
32 Kaijian He et al. / Procedia Computer Science 139 (2018) 25–32
8 K. HE, L. JI, G. TSO, B. ZHU, Y. ZOU / Procedia Computer Science 00 (2018) 000–000

[2] K. Chen, S. Zhang, What’s news in exchange rate dynamics: A {DSGE} approach, Economics Letters 134 (2015) 133 – 137.
[3] K. Dowd, Measuring Market Risk, John Wiley & Sons Inc., Chichester, England; Hoboken, NJ, 2005, 9780470013038; Kevin Dowd.; :ill. ;25
cm. +1 CD-ROM (4 3/4 in.; Includes bibliographical references (p. [365]-377) and index.
URL http://lib.cityu.edu.hk/record=b1977867;http://www.loc.gov/catdir/toc/ecip0511/2005010796.html
[4] R. A. Meese, K. Rogoff, Empirical exchange-rate models of the seventies - do they fit out of sample, Journal of International Economics
14 (1-2) (1983) 3–24.
[5] L. Yu, S. Wang, K. K. Lai, Forecasting crude oil price with an emd-based neural network ensemble learning paradigm, Energy Economics
30 (5) (2008) 2623–2635. doi:10.1016/j.eneco.2008.05.003.
[6] K. HE, Ensemble forecasting of value at risk via multi resolution analysis based methodology in metals markets, Working paper, City University
of Hong Kong (2010).
[7] K. He, C. Xie, S. Chen, K. K. Lai, Estimating var in crude oil market: A novel multi-scale non-linear ensemble approach incorporating wavelet
analysis and neural network, Neurocomputing 72 (16-18, Sp. Iss. SI) (2009) 3428–3438. doi:10.1016/j.neucom.2008.09.026.
[8] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-C. Yen, C. C. Tung, H. H. Liu, The empirical mode decomposition and
the hilbert spectrum for nonlinear and non-stationary time series analysis, Proceedings of the Royal Society of London. Series A: Mathematical,
Physical and Engineering Sciences 454 (1971) (1998) 903–995.
[9] N. E. Huang, M.-L. C. Wu, S. R. Long, S. S. P. Shen, W. Qu, P. Gloersen, K. L. Fan, A confidence limit for the empirical mode decomposition
and hilbert spectral analysis, Proceedings: Mathematical, Physical and Engineering Sciences 459 (2037) (2003) 2317–2345.
[10] A. Das, P. Das, Chaotic analysis of the foreign exchange rates, Applied Mathematics and Computation 185 (1) (2007) 388 – 396.
[11] P. Caraiani, E. Haven, Evidence of multifractality from {CEE} exchange rates against euro, Physica A: Statistical Mechanics and its Applica-
tions 419 (2015) 395 – 407.
[12] L. Yu, S. Y. Wang, K. K. Lai, A novel nonlinear ensemble forecasting model incorporating glar and ann for foreign exchange rates, Computers
& Operations Research 32 (10) (2005) 2523–2541.
[13] R. Hrasko, A. G. Pacheco, R. A. Krohling, Time series prediction using restricted boltzmann machines and backpropagation, Procedia Com-
puter Science 55 (2015) 990 – 999, 3rd International Conference on Information Technology and Quantitative Management, ITQM 2015.
doi:https://doi.org/10.1016/j.procs.2015.07.104.
[14] A. Dedinec, S. Filiposka, A. Dedinec, L. Kocarev, Deep belief network based electricity load forecasting: An analysis of macedonian case,
Energy 115, Part 3 (2016) 1688 – 1700, sustainable Development of Energy, Water and Environment Systems.
[15] C. Krauss, X. A. Do, N. Huck, Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the s&p 500, European
Journal of Operational Research 259 (2) (2017) 689 – 702. doi:https://doi.org/10.1016/j.ejor.2016.10.031.
[16] F. Shen, J. Chao, J. Zhao, Forecasting exchange rate using deep belief networks and conjugate gradient method, Neurocomputing 167 (2015)
243 – 253.
[17] G. E. Hinton, R. R. Salakhutdinov, Reducing the dimensionality of data with neural networks, Science 313 (5786) (2006) 504.
[18] Z. Zhang, Q. He, J. Gao, M. Ni, A deep learning approach for detecting traffic accidents from social media data, Transportation Research Part
C: Emerging Technologies 86 (2018) 580 – 596. doi:https://doi.org/10.1016/j.trc.2017.11.027.
[19] D. Y. Kim, H. Y. Song, Method of predicting human mobility patterns using deep learning, Neurocomputing 280 (2018) 56 – 64, applications
of Neural Modeling in the new era for data and IT. doi:https://doi.org/10.1016/j.neucom.2017.07.069.
[20] J. Qiao, G. Wang, W. Li, X. Li, A deep belief network with plsr for nonlinear system modeling, Neural Networks 104 (2018) 68 – 79.
doi:https://doi.org/10.1016/j.neunet.2017.10.006.
[21] C. Luo, D. Wu, D. Wu, A deep learning approach for credit scoring using credit default swaps, Engineering Applications of Artificial Intelli-
gence 65 (2017) 465 – 470. doi:https://doi.org/10.1016/j.engappai.2016.12.002.
[22] G. Fu, Deep belief network based ensemble approach for cooling load forecasting of air-conditioning system, Energy 148 (2018) 269 – 282.
doi:https://doi.org/10.1016/j.energy.2018.01.180.
[23] L. Yu, R. Zhou, L. Tang, R. Chen, A dbn-based resampling svm ensemble learning paradigm for credit classification with imbalanced data,
Applied Soft Computing 69 (2018) 192 – 202. doi:https://doi.org/10.1016/j.asoc.2018.04.049.
[24] M. H. Rafiei, H. Adeli, A novel machine learning model for estimation of sale prices of real estate units, Journal of Construction Engineering
and Management 142 (2) (2016) 04015066.
[25] Q. Zhang, L. T. Yang, Z. Chen, P. Li, A survey on deep learning for big data, Information Fusion 42 (2018) 146 – 157. doi:https:
//doi.org/10.1016/j.inffus.2017.10.006.
[26] C. Brooks, Introductory Econometrics for Finance, 2nd Edition, Cambridge [England] ; New York : Cambridge University Press, 2008.
[27] Quandl, EIA and OPEC various end-of-day data (2018).
URL https://www.quandl.com/data/EIA-U-S-Energy-Information-Administration-Data,https://www.quandl.com/
data/OPEC/ORB-OPEC-Crude-Oil-Price
[28] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow,
A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané, R. Monga, S. Moore, D. Murray, C. Olah,
M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Viégas, O. Vinyals, P. Warden, M. Watten-
berg, M. Wicke, Y. Yu, X. Zheng, TensorFlow: Large-scale machine learning on heterogeneous systems, software available from tensorflow.org
(2015).
URL https://www.tensorflow.org/
[29] albertbup, A python implementation of deep belief networks built upon numpy and tensorflow with scikit-learn compatibility (2017).
URL https://github.com/albertbup/deep-belief-network