Deep Learning for Predictions in Emerging Currency Markets
Svitlana Galeshchuk
1,2
and Sumitra Mukherjee
3
1
Department of Accounting and Audit, Ternopil National Economic University, Ternopil, Ukraine
2
Laboratoire d'Informatique de Grenoble, Université Grenoble Alpes, Grenoble, France
3
College of Engineering and Computing, Nova Southeastern University, Fort Lauderdale, U.S.A.
Keywords: Neural Networks, Deep Learning, Convolution Networks, Exchange Rate Prediction, Emerging Markets.
Abstract: Accurate prediction of exchange rates is critical for devising robust monetary policies. Machine learning
methods such as shallow neural networks have higher predictive accuracy than time series models when
trained on input features carefully crafted by domain knowledge experts. This suggests that deep neural
networks, with their ability to learn abstract features from raw data, may provide improved predictive
accuracy with raw exchange rates as inputs. The preponderance of research focuses on developed currency
markets. The paucity of research in emerging currency markets, and the crucial role that stable currencies
play in such economies, motivates us to investigate the effectiveness of deep networks for exchange rate
prediction in emerging markets. Literature suggests that the Efficient Market Hypothesis, which posits that
asset prices reflect all relevant information, may not hold in such markets because of extraneous factors
such as political instability and governmental interventions. This motivates our hypothesis that inclusion of
carefully chosen macroeconomic factors as input features may improve the predictive accuracy of deep
networks in emerging currency markets. This position paper proposes novel input features based on
currency clusters and presents our method for investigating the hypothesis using exchange rates from
developed as well as emerging currency markets.
1 INTRODUCTION
Transactions worth billions of dollars a day take
place in the foreign exchange market, making it one
of the largest financial markets in the world (Report
on global foreign exchange market activity in 2013).
Exchange rates are expressed in terms of a base-
quote currency pair that represents the number of
units of quote currency that may be exchanged for
each unit of the base currency. Accurate prediction
of forex rate rates is critical for formulating robust
monetary policies and developing effective trading
and hedging strategies in the foreign exchange
market (Lukas and Taylor, 2007)
Econometric models are not effective for
exchange rate predictions when the forecast horizon
is less than a year (Meese and Rogoff, 1983). Time
series models are poor at predicting the direction of
change in rates. Shallow artificial neural networks
and support vector machines perform marginally
better when using carefully crafted input features;
significant efforts by domain experts may be needed
to obtain such features from raw input data.
The recent success of deep neural networks in a
variety of domains may be partially attributable to
their ability to learn abstract features from raw data
(LeCun et al., 2015). This suggests that deep
networks may be effective in predicting foreign
exchange rates based on raw time series data.
Our first objective is to investigate whether deep
neural networks are significantly better at foreign
exchange rate prediction than time series models and
shallow networks when raw exchange rate data are
used as input features. Our preliminary results using
exchange rates between the US dollar and three
major currencies in mature markets–Euro, British
Pound, and Japanese Yen–suggest that indeed deep
convolution networks perform better than extant
methods.
The preponderance of research in foreign
exchange prediction focuses on established markets.
In response to the paucity of research in emerging
currency markets, and in recognition of the fact that
stable currency markets play a crucial role in
determining the well-being of such economies, our
second objective is to adapt deep network models
for predicting exchange rates in emerging markets.
Galeshchuk S. and Mukherjee S.
Deep Learning for Predictions in Emerging Currency Markets.
DOI: 10.5220/0006250506810686
In Proceedings of the 9th International Conference on Agents and Artificial Intelligence (ICAART 2017), pages 681-686
ISBN: 978-989-758-220-2
Copyright
c
2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
681
As representative emerging markets we consider
countries in the Eastern Partnership (EaP). The
Eastern Partnership is an initiative of the European
Union that aims to foster improved economic
relationship with the post-Soviet states of Armenia,
Azerbaijan, Belarus, Georgia, Moldova, and
Ukraine. Improved macroeconomic conditions in the
EaP countries is a pre-condition for their economic
integration with European Union. Research suggests
that currency market stability is one of the most
important indicators of sustainable development and
growth in these economies and that accurate
prediction of exchange rate is critical to the
formulation of robust monetary policies. This lends
further impetus to our study of developing improved
models for exchange rate prediction in emerging
markets.
Literature suggests that the Efficient Market
Hypothesis, which posits that asset prices reflect all
relevant information, may not hold in emerging
markets because of extraneous factors such as
political instability and governmental interventions.
This motivates our hypothesis that inclusion of
carefully chosen macroeconomic factors as input
features may improve the predictive accuracy of
deep networks in emerging currency markets. An
ancillary goal of this study is to develop a novel set
of input features that are obtained by forming
clusters of currency markets based on distance
metrics derived from correlation measures.
The roadmap for the remainder of this position
paper is as follows: Section 2 formally defines the
exchange rate prediction problem. Section 3 briefly
discusses the related literature. Section 4 describes
our proposed methodology. Section 5 concludes
with some observations.
2 THE PREDICTION PROBLEM
We use a standard formulation of the exchange rate
prediction problem where our goal is to predict the
direction of change: Let
and
denote the
values of an exchange rate between a pair of
currencies in periods and +, respectively, for
some >0. Define the direction of change
(
)
=
1 if the rate increases in periods, i.e. if
−
>0; otherwise,
(
)
=0. Our objective is to
learn a function
:ℝ
→
0,1
such that
,
,…,
=
(
)
. We train models to
predict the direction of change. Let ̂
(
)
=
,
,…,
be the predicted direction of
change periods forward, where
is a function
learnt by a model. A period forward prediction
model model is evaluated by its classification
accuracy on out-of-sample observations, where
classification accuracy is defined as the percentage
of test cases for which the predicted direction of
change ̂
(
)
equals the true direction of change
(
)
.
3 RELATED WORK
Exchange rate prediction methods may be
categorized into econometric methods, time series
models, and machine learning techniques. We re-
view these approaches briefly and then discuss deep
neural networks.
3.1 Econometric Models
Econometric models predict exchange rate based on
economic factors. The Mundell-Fleming model
(1962), Dornbusch’s (1976) asset-market approach
to exchange-rates, and New Keynesian models are
examples of such models and a good survey of such
models can be found in Engel (2013). These models
are widely used by central bankers around the world.
However, research indicates that these models are
not effective when the prediction horizon is less than
a year (Neely and Sarno, 2002).
Meese and Rogoff (1983) demonstrated that such
models fail to outperform a random walk in out of
sample predictions and their findings are still widely
accepted.
3.2 Time Series Models
An excellent survey of time series forecasting
models can be found in Box et al. (2015).
Autoregressive Integrated Moving Average
(ARIMA) models and Exponential Smoothing (ETS)
models are the most commonly used time series
models for foreign exchange rate prediction.
ARIMA models can deal with non-stationary data by
differencing transformations and subsume
autoregressive models and moving average models
as special cases. ETS models are non-stationary and
can capture trends and seasonality. Time series
models may provide satisfactory point estimates for
exchange rates, but the direction of change implied
by these estimates are often poor indicators of the
true direction.
ICAART 2017 - 9th International Conference on Agents and Artificial Intelligence
682
3.3 Artificial Neural Networks
Artificial neural network (ANN) with a single
hidden layer often outperform time series models in
providing point estimates for exchange rates as
demonstrated in Dunis (2015) Thinyane and Millin
(2011), Nag (2002), and Galeshchuk (2016).
However, the direction of change implied by these
point estimates are often unacceptably inaccurate.
This renders these method less useful as a basis for
formulating monetary policies. This further
motivates us to investigate the ability of deep
networks to predict the direction of change in forex
rates.
3.4 Deep Neural Networks
Deep learning techniques originally introduced by
Ivakhnenko (1971) and then Hinton (2002, 2006)
has been successfully applied in a variety of
domains including face detection (Osadchy et al.,
2013), speech recognition (Sukittanon et al., 2004),
object recognition (Schmidhuber, 2005), document
categorization (Hinton and Salakhutdinov, 2006),
and natural language processing (Lee et al., 2009).
Deep learning networks have also been used for time
series predictions (Busseti et al., 2012; Langkvist et
al., 2014) and for financial predictions (Ribeiro and
Noel, 2011; Chao et al. 2011; Yeh et al., 2014; Lai et
al.). Restricted Boltzmann machines and auto-
encoders machines have been used for
dimensionality reduction and unsupervised pre-
training. Applications are discussed in Larochelle et
al. (2009), Masci et al. (2011), and Vincent et al.
(2007).
Deep convolution networks (DN) are attractive
for high dimensional prediction and classification
problems (LeCun et al 2015). DNs are suitable for
exchange rate prediction for two main reasons: First,
high level features abstracted by the network may
serve as noise filters and dimensionality reduction
techniques may help abstract input features.
Secondly, the temporally-local correlation between
consecutive observations may be exploited to reduce
the number of parameters to be estimated in the
network by connecting only a small number of
adjacent inputs to each unit in a hidden layer.
Our work is motivated by results from
experiments to compare the accuracy of deep
networks with baseline models (ARIMA, ETS, and
ANN) to predict the direction of changes of
exchange rates for EUR/USD, GBP/USD, and
USD/JPY (Galeshchuk and Mukherjee, 2017).
Results demonstrate that trained deep networks
achieve better out-of-sample prediction accuracy
than baseline methods.
Units in a DN receive inputs from small
contiguous receptive fields that collectively cover
the entire set of input features. This allows units to
act as local filters and to exploit local correlation
between contiguous inputs. Units share weights and
bias parameters to create a feature map and this not
only results in a significant reduction in the number
of parameters to be estimated but also facilitates
detection of features irrespectively of their actual
position in the input field. The reduction in the
number of parameters may be very significant as the
number layers in the network and the number of
units in each layer increases.
Recurrent neural networks are an effective class
of neural network designed to handle sequence
dependence. Stacked Long Short-Term Memory
(LSTM) is a type of recurrent neural network used in
deep learning which makes effective use of model
parameters, converges quickly, and outperforms
deep feed forward neural networks. That is why, it is
often used for time-series predictions. Being adapted
for dimensionality reduction and unsupervised pre-
training tasks, LSTMs have been successfully used
for unsupervised extraction of abstract input features
for prediction problems. The approach has also
proved effective in financial predictions.
4 METHODOLOGY
In this section we describe the data sets to be used in
this study, discuss additional features to be used for
prediction in emerging markets, present baseline
models including shallow neural networks, and
describe our deep convolution networks.
4.1 Data Sets
For developed currency markets, we use the daily
closing rates between three currency pairs: Euro and
US Dollar (EUR/USD), British Pound and US
Dollar (GBP/USD), and US Dollar and Japanese
Yen (USD/JPY) to train and test our models. The
rates may be downloaded from: http://www.global-
view.com/forex-trading-tools/forex-history/. Data
for the years 2000 to 2015 are considered. For
emerging currency markets we use the exchange
rates of EaP countries to US Dollar: AZN/USD,
AMD/USD, BYR/USD, MDL/USD, UAH/USD,
GEL/USD. For each data set we train models for
daily, monthly, and quarterly predictions.
Deep Learning for Predictions in Emerging Currency Markets
683
4.2 Input Macroeconomic Features
In order to provide better exchange-rates prediction
on the macroeconomic level, researchers develop
monetary models of exchange rates based on
fundamental economic data. We will include the
indicators of real sector (GDP growth,
unemployment, wages), current and capital account
(current account balance, openness as ratio of total
import and export to GDP), public and private
foreign debt, capital flows, and ratio of international
reserves to 3 months import, international variables
(interest rates and price ratios). Some additional
factors that may need to be considered include:
money growth, fiscal growth, and a measure for the
degree of political instability and market
liberalization.
Improved exchange rate prediction models are
particularly challenging to develop in volatile
emerging markets with political instability as is the
case in EaP economies. The EU is the main
economic partners of EaP states. Financial markets
of EaP countries and Russia are still highly coupled
through trade and political relationships in post-
soviet period. The high co-volatility of these markets
requires us to identify distinct patterns of linkages
among European, EaP, and Russian markets.
Furthermore, contagious effect of crises is observed
widely as local currency deterioration worsens
macroeconomic indicators in trading partners.
The core currencies in EU-EaP-Russia area will
be modelled as a network. The correlation between
these exchange rates will be computed for a selected
time horizon. We will use a 3 month horizon since
international trading the payments are made up to 90
days. Then, each correlation coefficient in the
correlation matrix of the N markets will be mapped
to a metric distance between pairs of indices to form
an N×N distance matrix with values ranging
between 0 and 1. This distance matrix will be used
to construct a minimal spanning tree (MST) in a
fully connected graph where the vertices represent
the currencies and the arc lengths inversely
proportion to the strength of the correlations
between the currencies. Clusters will be formed by
removing the longest edges of the MST. Strongly
correlated currencies are connected by short links
and belong to the same cluster; unrelated currencies
connected by longer links belong to different
clusters. This will provide insights regarding the
pattern of currency crises spread in the EaP
economies and permit us to investigate
synchronization among the currency markets in the
EaP area.
4.3 Baseline Models
We use a random walk model, two time series
models (ARIMA and ETS), and a single layered
neural network as baseline models. The time series
models provide point estimates
for the rates.
We predict output class ̂
(
)
=1 if
>
, and
0 otherwise. The predicted direction of change ̂
(
)
is compared with the actual direction of change
(
)
. Results for ARIMA and ETS are obtained
using the auto.arima model and the ets model from
the R library forecast with default parameters
(Hyndman and Khandakar 2008).
A neural network model with a single hidden
layer will also be used in our study as a baseline
model. The units have sigmoid transfer functions
and use gradient descent and backpropagation for
training. The model is trained on vectors with
features
,
,…,
as inputs and
as
output to predict a point estimate
for the
period forward rate. As in the case of the time series
models, we predict the output class ̂
(
)
=1 if
>
, and 0 otherwise to compare the actual
and predicted directions of change. Results are
obtained using the R package nnet. Models
parameters are tuned through cross-validation by
performing a grid search over the parameter ranges
using the tune function from the R package e1071.
For details of these packages, see https://cran.r-
project.org/web/packages/nnet/nnet.pdf and: https://
cran.r-project.org/web/packages/e1071/e1071.pdf).
4.4 Deep Convolution Network
The deep convolution network has layers of hidden
units separating the input layer from the output unit.
We use
to denote the internal bias of the
th
unit in
the
th
layer and
to represent the weight of the
connection to that unit from the
th
unit in the
(
−1
)
th
layer. For an input vector , the output of
th
unit in the
th
layer is computed as ℎ
(
)
=
, where
=
+
∑
ℎ
(
)
, and
(
)
=max(0,) is the rectified linear unit
function. The output uses a softmax transfer
function. Adam optimizer (Kingma et al 2015) is
used to minimize a cross-entropy loss function. The
open source library TensorFlow is used to create the
DN models (https://www.tensorflow.org/).
4.5 Stacked Long Short-term Memory
We intend to use Stacked Long Short-Term Memory
(LSTM) deep network with mechanisms for
ICAART 2017 - 9th International Conference on Agents and Artificial Intelligence
684
exchange-rate prediction in this experiment. LSTM
network is a type of recurrent neural network used in
deep learning because very large architectures can
be successfully trained.
The output value of recurrent neural network
(Galeshchuk, 2014) can be formulated as:
=
(
∑
−
),
ℎ
=
(
+
(
−1
)
+3
(
−1
)
−
)
where
,
are logistic activation functions, is the
number of neurons in the hidden layer,
is the
weight coefficient from -neuron of the hidden layer
to the output neuron,
is the output value of -
neuron of the hidden layer,
is the threshold of the
output neuron, is the number of neurons in the
input layer,
are the weight coefficients from the
i -input neuron to -neuron of the hidden layer,
are the input values,
are the thresholds of the
neurons of the hidden layer,
is the synapse from
context neuron of the hidden layer to the -neuron
of the same (hidden) layer,
(
−1
)
is the output
value of context neuron of hidden layer in the
previous moment of time−1,3is the synapse
from context output neuron to the -neuron of the
hidden layer, (−1)is the value of context output
neuron in the previous moment of time
1t
For the version of LSTM used, is implemented
by the following composite function (see Graves at
al., 2013):
=(
+
ℎ
+
+
)
=(
+
ℎ
+
+
)
=
+
tanh(
+
ℎ
+
)
=(
+
ℎ
+
+
)
ℎ
=
tanh(
)
where is the logistic sigmoid function, and ,,,
are respectively the input gate, forget gate, output
gate and cell activation vectors, all of which are the
same size as the hidden vector ℎ.
5 CONCLUSIONS
This position paper outlines our approach for
developing improved models for exchange rate
prediction using deep neural networks. The ability of
deep networks to learn abstract features from raw
data motivates this approach. Preliminary results
confirm that our deep network produces
significantly higher predictive accuracy than the
baseline models for developed currency markets. We
now plan to adapt this model for exchange rate
prediction in emerging currency markets by
including macroeconomic factors as input features.
A novel set of input features based on currency
clusters may help improve predictive accuracy of
such models. This study will be among the first to
integrate information about market liberalization and
political stability with macroeconomic indicators
and time-series data on exchange rate and
transaction volume. Inclusion of these factors as
predictors should improve predictive accuracy for
exchange rate, especially in emerging markets.
REFERENCES
Box G. E. P., Jenkins G. M., Reinsel G. C., Ljung G. M.
2015. Time Series Analysis: Forecasting and Control,
5th Edition, Wiley.
Busseti E., Osband I., Wong S. 2012. Deep Learning for
Time Series Modeling. CS 229 Final Project Report.
Chao J., Shen F., Zhao J. 2011. Forecasting Exchange
Rate with Deep Belief Networks. Proceedings of
International Joint Conference on Neural Networks,
San Jose, California, USA.
Dornbusch R. 1976. Exchange Rate Expectations and
Monetary Policy. Journal of International Economics
6 (3): 231–244.
Dunis C. L., Laws J., Sermpinis G. 2011. Higher order and
recurrent neural architectures for trading the
EUR/USD exchange rate. Quantitative Finance 11(4):
615-629.
Engel. C. 2013. Exchange rates and interest parity.
National Bureau of Economic Research: 77.
Fleming J. M. 1962. Domestic financial policies under
fixed and floating exchange rates. IMF Staff Papers 9:
369–379.
Galeshchuk S. 2016. Neural networks performance in
exchange rate prediction. Neurocomputing 172: 446-
452.
Galeshchuk, S., 2014. Neural-based method of measuring
exchange-rate impact on international companies’
revenue. In Distributed Computing and Artificial
Intelligence, 11th International Conference. Springer
International Publishing: 529-536.
Galeshchuk, S., Mukherjee S., 2016 Deep Networks for
Predicting Direction of Change in Foreign Exchange
Rates. Intelligent Systems in Accounting, Finance and
Management: early view papers.
Graves, A., Mohamed, A.R. and Hinton, G., 2013, May.
Speech recognition with deep recurrent neural
networks. In 2013 IEEE international conference on
acoustics, speech and signal processing (pp. 6645-
6649). IEEE.
Hinton G. E. 2002. Training products of experts by
minimizing contrastive divergence. Neural Comput.
14: 1771–1800.
Hinton G. E., Osindero S., The Y. 2006. A fast learning
Deep Learning for Predictions in Emerging Currency Markets
685
algorithm for deep belief nets. Neural Computations.
18: 1527–1554.
Hinton G. E., Salakhutdinov R. 2006.Reducing the
dimensionality of data with neural networks. Science.
313 (5786): 504–507.
Hyndman R. J., Khandakar Y. 2008. Automatic time
series forecasting: the forecast package for R,. Journal
of Statistical Software 26 (3): 1-22, 2008 DOI:
http://ideas.repec.org/a/jss/jstsof/27i03.html.
Kingma, D. P., Ba, J. L. (2015). Adam: a Method for
Stochastic Optimization. International Conference on
Learning Representations, 1–13.
Lai A., Li M. K., Pong F.W. Forecasting Trade Direction
and Size of Future Contracts Using Deep Belief
Network. Stanford University.
Langkvist M., Karlsson L., A. Loutfi. 2014. A review of
unsupervised feature learning and deep learning for
time-series modeling. Pattern Recognition Letters 42:
11–24.
Larochelle H., Bengio Y., Louradour, P. Lamblin. 2009.
Exploring strategies for training deep neural networks.
The Journal of Machine Learning Research 10: 1-40.
LeCun Y., Bengio Y., Hinton G. 2015. Deep Learning.
Nature 521: 436–444.
LeCun Y., Bottou L., Bengio Y., Haffner P. 1998.
Gradient-based learning applied to document
recognition. Proc. IEEE. 86(11); 2278–2324.
Lee H., Largman Y., Pham P., Ng A. 2009. Unsupervised
feature learning for audio classification using
convolutional deep belief networks. Advances in
Neural Information Processing Systems 22.
Lukas M., Taylor M. 2007. The Obstinate Passion of
Foreign Exchange Professionals: Technical Analysis.
Journal of Economic Literature 45 (4): 936–972.
Masci J., Meier U., Ciresan D., Schmidhuber J. Stacked
Convolutional Auto-Encoders for Hierarchical Feature
Extraction. Lecture Notes in Computer Science 6791:
52-59.
Meese R., Rogoff K. 1983. The Out-of-Sample Failure of
Empirical Exchange Rate Models: Sampling Error or
Misspecification? NBER Chapters, in Exchange Rates
and International Macroeconomics: pp. 67–112.
Mundell R. A. 1963. Capital mobility and stabilization
policy under fixed and flexible exchange rates.
Canadian Journal of Economic and Political Science
29 (4): 475–485.
Nag A. 2002. Forecasting daily foreign exchange rates
using genetically optimized neural networks. Journal
of Forecasting 21(7), pp. 501- 511, 2002.
Neely C., Sarno L. 2002. How well do monetary
fundamentals forecast exchange rates? Federal
Reserve Bank of St. Louis Working Paper Series:
2002-2007.
Osadchy M., LeCun Y., Miller M. 2013. Synergistic face
detection and pose estimation with energybased
models. Journal of Machine Learning Research 8
:
1197–1215.
Report on global foreign exchange market activity in
2013. April 2013. Triennial Central Bank Survey.
Basel, Switzerland: Bank for International
Settlements. http://www.bis.org/publ/rpfx13fx.pdf.
Ribeiro B., Noel L. Deep Belief Networks for Financial
Prediction. Proceedings of ICONIP 2011, Part III,
LNCS 7064; 766–773.
Schmidhuber J. 2005. Deep Learning in Neural Networks:
An Overview. Neural Networks: 85-117.
Simard P. Y., Steinkraus D., Platt J. C.,. 2003. Best
Sukittanon S., Surendran A.C., Platt J. C., Burges C. J.
2004. Convolutional networks for speech detection.
Interspeech: 1077–1080.
Thinyane H., Millin J. 2011. An investigation into the use
of intelligent systems for currency trading.
Computational Economics 37(4): 363-374.
Vincent P., Larochelle H., Bengio Y., Manzagol P.
Extracting and Composing Robust Features with
Denoising Autoencoders,” Proceedings of the 25th
International Conference on Machine Learning (ICML
08); 1096-1103.
Wagner N., Michalewicz Z., Khouja M., McGregor R. R.
2007. Time Series Forecasting for Dynamic
Environments: The DyFor Genetic Program Model.
Trans. Evol. Comp. 11 (4): 433-452.
Xiao R. 2014. Deepnet: deep learning toolkit in R. R
package version 0.2. http://CRAN.R-
project.org/package=deepnet.
Yeh S-H., Wang C.J., Tsai M.F. 2014. Corporate Default
Prediction via Deep Learning. ISF.
ICAART 2017 - 9th International Conference on Agents and Artificial Intelligence
686
