introduction - journal€¦ · web viewtrade implications of new growth theories are that trade...
TRANSCRIPT
Cover Letter
Cover Letter
Dear Sir/Ma’am,
I am Manas Tripathi, faculty at Indian Institute of Management (IIM) Rohtak, India. I along with my other co-authors Saurabh Kumar (Faculty at IIM Indore), and Sarveswar kumar Inani (Faculty at Jindal Global Business School) have worked on a novel problem of Exchange Rate Forecasting Using Ensemble Modeling for Better Policy Implications. The study have implications for policymakers, regulators, investors, speculators, and arbitrageurs.
We have used data of 7 years for three of the most traded currency pairs (EUR/USD, GBP/USD, and JPY/USD) and forecasted the next value using linear, non-linear, hybrid and ensemble methods. The highlights of the manuscript are given below:
1. An ensemble technique is proposed to forecast daily exchange rates for three currency pairs.
2. Ensemble combines three models: mean forecast, ARIMA, and artificial neural network.
3. Proposed methodology is able to forecast better as compared to individual models.
4. The study has implications for policymakers, regulators, investors, and speculators.
I am submitting the current manuscript for its possible publication in the prestigious Journal of Policy Modeling. I confirm that the manuscript is the original research article and has not been submitted elsewhere in any of the conferences or journals.
I hope the current manuscript aligns with the journal scope and objectives. Please let me know if anything else is required from my end.
Thanks and Regards
Manas Tripathi
1
Title Page
Exchange Rate Forecasting Using Ensemble Modeling for Better Policy Implications
Manas Tripathia (Corresponding Author)a Faculty, Management Information Systems Area, Indian Institute of Management Rohtak, India
Email: [email protected]
Saurabh Kumarb
b Faculty, Information Systems Area, Indian Institute of Management Indore, India
Address: C-Block, Indian Institute of Management Indore, Indore, India- 453556
Phone No: +91-8542091617, Email: [email protected]
Sarveshwar Kumar Inanic
c Faculty, Finance Area, Jindal Global Business School, India
Email: [email protected]
2
Anonymous Manuscript
Exchange Rate Forecasting Using Ensemble Modeling for Better Policy Implications
Abstract: This study aims to contribute in the area of foreign exchange forecasting. Exchange
rate plays an essential role for the economic policy of a country. Due to the floating exchange
rate regime, and ever-changing economic conditions, analysts have observed significant
volatility in the exchange rates. However, exchange rate forecasting has been a challenging task
before the analysts over the years. Various stakeholders such as the central bank, government,
and investors try to maximize the returns and minimize the risk in their decision-making using
exchange rate forecasting. The study aims to propose a novel ensemble technique to forecast
daily exchange rates for the three most traded currency pairs (EUR/USD, GBP/USD, and
JPY/USD). The ensemble technique combines the linear and non-linear time-series forecasting
techniques (mean forecast, ARIMA, and neural network) with their most optimal weights. We
have taken the data of more than seven years, and the results indicate that the proposed
methodology could be an effective technique to forecast better as compared to the component
models separately. The study has crucial economic and academic implications. The results
derived from this study would be useful for policymakers, regulators, investors, speculators, and
arbitrageurs.
Keywords: Neural Network, Currency Pairs, Forecasting, Ensemble, ARIMA Models, Exchange
Rate
JEL Classification: G15, G17, F31, C45
3
Anonymous Manuscript
1 Introduction
Foreign exchange rates are one of the very important factors in the international monetary
markets. It is widely known that exchange rates are affected by many micro-, macro-economic,
and political factors. This has motivated the researchers and practitioners to find the good
explanation for the movement of exchange rates (Ince & Trafalis, 2006; Qureshi, Rehman, &
Qureshi, 2018). Exchange rate forecasting has been an active area of research in computational
intelligence and econometrics. Exchange rate forecasting is an important constituent for the
economies especially in the context of leading economies as these economies are very sensitive
to external negative news. Researchers have proposed various econometric and time series
models to understand the volatility of exchange rates.
Time series forecasting has received significant attention across the world, in different
domains of management research, due to its applicability in different areas. One of the most
crucial areas in time series forecasting is modeling and forecasting a financial time series, such
as stock index, individual stocks, exchange-traded funds, exchange rates, etc. These financial
time series are affected by several economic and political conditions which include the stability
of the government, interest and inflation rates, national output, growth rate, employment, etc.
Therefore, it becomes imperative to identify an adequate model or underlying data generating
process, which can explain a financial time series. Time series forecasting is very useful
whenever there is very little information available about possible explanatory variables. Hence,
investors use technical analysis to forecast a financial time series. Technical analysis relies on
the assumption that historical patterns of a series could be used to forecast short-term and
medium-term forecasts of that series. In the long-run, technical analysis does not work well
because a financial time series depends on other economic, political, social, and psychological
4
Anonymous Manuscript
factors. In the literature, many studies have focused on the modeling and forecasting of a
financial time series (Adhikari & Agrawal, 2014; Buckley & O’Brien, 2017; Hussain, Knowles,
Lisboa, & El-Deredy, 2008; Prusa, Sagul, & Khoshgoftaar, 2018; Rodríguez-González, Colomo-
Palacios, Guldris-Iglesias, Gómez-Berbís, & García-Crespo, 2012).
There are various linear and non-linear methods in the literature to model and forecast a time
series. Among the linear models, Autoregressive Integrated Moving Average (ARIMA) (Box &
Jenkins, 1976) models have become a standard technique. The ARIMA model is discussed in
detail in the methodology section. ARIMA models without any autoregressive and moving
average term are called random walk model. Random walk models have found extensive use in
forecasting a financial time series, particularly exchange rates (Zhang, 2003). Though linear
models are simple and easy to use, the main problem is that they are not able to capture any
nonlinearity associated with the data (Dunis, Laws, & Schilling, 2012; Zhang, 2003).
Nonlinearity is an inherent characteristic of any financial time series because it gets affected by
many economic and social factors. Balkin and Ord (2000) suggest that artificial neural networks
(ANNs) are universal approximators and they are capable of identifying any nonlinear
relationships in the data. Therefore, ANNs could provide a promising solution to capture such
nonlinearities. The ANNs have found massive applications in the forecasting of economic and
financial time series (Alfaro, García, Gámez, & Elizondo, 2008; Choi, Yu, & Au, 2011; Huang,
Chen, Hsu, Chen, & Wu, 2004; Khansa & Liginlal, 2011; Lam, 2004; Lee & Yum, 1998; Pasley
& Austin, 2004; Panda & Narasimhan, 2007). ANNs have gained such popularity because of
their nonlinear, self-adaptive, data-driven, and non-parametric properties (Khashei & Bijari,
2010, 2011). Because of these salient characteristics, ANNs are proven to be better in forecasting
5
Anonymous Manuscript
financial time series having non-linearity. The extensive use of neural networks in various
business applications is shown in the systematic review by (Tkáč & Verner, 2016).
The studies comparing the performance of ARIMA and ANN models have yielded mixed
and inconsistent results (Zhang, Eddy Patuwo, & Y. Hu, 1998). Some studies have established
that ANNs forecast better as compared to ARIMA and random walk models (Dunis et al., 2012;
Ghazali, Jaafar Hussain, Mohd Nawi, & Mohamad, 2009; Sermpinis, Dunis, Laws, & Stasinakis,
2012). Whereas, some other studies have found that the forecasting accuracies of ANN are worse
than the simple random walk model for exchange rate forecasting (Hann & Steurer, 1996;
Taskaya-Temizel & Casey, 2005). Therefore, none of the methodologies is superior in all the
situations, and the blind use of any of them could provide inaccurate forecasting results. The
literature has confirmed that a financial time series contains both linear and non-linear patterns,
and hence, ANNs and ARIMA models could be used together to get better forecasting results
leading to better accuracy (Khashei & Bijari, 2011; Zhang, 2003). Therefore, we move towards
models, which combine random walk or ARIMA and ANN models. Such hybridization of
ARIMA and ANN was first proposed by Zhang (2003) and then popularized by some other
studies (Adhikari & Agrawal, 2014; Khashei & Bijari, 2010, 2011). The hybrid models are based
on the assumption that a financial time series contains linear and non-linear components, which
could be modelled separately from the series. In such models, first, the linear part is modelled by
the random walk or ARIMA models then the residuals of the model are computed. After
capturing the linear relationship, these residuals contain only nonlinear relationship. Then, ANN
models are used to model these residuals, and then the forecasts for the financial time series are
generated. The problem arises here because the forecasts of ARIMA and ANN models are
merely added, but their weights in the hybrid model are not considered to identify an optimal
6
Anonymous Manuscript
hybrid model. Therefore, we propose an ensemble model which calculate the optimum weights
for component models. These methods of combining models to form the ensemble are described
in detail in the methodology section.
We select exchange rate time series data because exchange rate forecasting has significant
implications for importers, exporters, investors, the central bank, and the government. We have
used the exchange rate data of the three most traded currency pairs (EUR/USD, GBP/USD, and
JPY/USD) to check the forecasting accuracy of our proposed ensemble model. We have taken
these currency pairs because these currencies belong to the most developed nations whose
economies impact the business of the entire world.
The motivation behind carrying out this study lies in the classical theories related to
international trade. The exchange rate fluctuation could impact international trade. The
contribution of the study is twofold. First, it proposes a novel ensemble technique to forecast
daily exchange rates for the three most traded currency pairs (EUR/USD, GBP/USD, and
JPY/USD). The ensemble technique combines the linear and non-linear time-series forecasting
techniques (mean forecast, ARIMA, and neural network) with their most optimal weights.
Second, the results of the study indicate that the proposed ensemble method could be an effective
technique to forecast better as compared to the component models separately. A hypothetical
case in the study also validates the predictive power of the ensemble model. The study has
crucial economic and academic implications. The results of the study would certainly be useful
for policymakers, regulators, investors, speculators, and arbitrageurs.
The paper consists of nine sections. Section 2 presents the theoretical foundation and
existing literature; Section 3 elaborates on the data collection; Section 4 illustrates the
methodology used in the study; Section 5 exhibits the empirical results; Section 6 discusses the
7
Anonymous Manuscript
results with a hypothetical case; Section 7 and Section 8 reveal the economic and academic
implications of the study; and finally, Section 9 concludes the study with limitations and the
future scope.
2 Theoretical Foundation
The motivation of this study lies in the heart of the theory of international trade, and theory
of trade and growth. Since exchange rate volatility and international trade are closely related, it is
pertinent to outline the theoretical background.
2.1 Theory of International Trade
The world has become a global market where countries exchange goods and services
amongst them. Countries differ in their productive capacity of goods and services, which forms
the basis for international trade. This difference in productive capacity causes differences in
prices, which are the main cause of international trade. International trade gives the opportunity
to every country to specialize in the production of those things in which they have a competitive
advantage. This results in the division of labor which enables the countries to focus on their
specialized skills and take advantage of the capabilities of other nations. Division of labor leads
to economies of scale which in turn leads to products with reduced cost and improved quality.
Authors have given different views on international trade. Adam Smith has described that
international trade takes place due to the absolute advantage countries have over other countries.
On the other hand, the Ricardian model suggests that the difference in technology and natural
resources are the basis for the countries to have a comparative advantage over other countries
(Dornbusch, Fischer, & Samuelson, 1977). The Ricardian model does not consider factor
endowment such as land, labor, and capital within a country. Alternatively, Heckscher–Ohlin
model (H–O model) proposes that differences in factor endowments are the basis for
8
Anonymous Manuscript
international trade between two countries (Heckscher & Ohlin, 1991). They further explain, that
a country exports those goods which are made by locally abundant resources in that country.
Similarly, a country will import those goods which require locally scarce resources.
The exchange rate is the price of one country’s currency with respect to the currency of
other country. Exchange rate, in principle, can affect international trade in many ways. Authors
have tried to understand the relationship between the volatility of the exchange rate and
international trade. Some authors have found that exchange rate volatility has an adverse effect
on international trade (Baron, 1976; Clark, 1973; Cushman, 1983). Péridy (2003) has found
mostly negative impact of exchange rate volatility on the export of G-7 countries. The study
further suggests that this impact of exchange rate volatility on exports is dependent on the
industries covered, and the destination market. Similarly, researchers have examined the impact
of exchange rate volatility on bilateral US trade using sectoral data (Byrne, Darby, &
MacDonald, 2008). They have found a significantly negative impact of exchange rate volatility
on US trade across sectors, strongest for exports of differentiated goods than that of homogenous
goods. However, authors have found a positive relationship between exchange rate volatility and
exports if the firms are able to reallocate their products to the domestic or foreign markets (Broll
& Eckwert, 1999). To deal with the fluctuations in exchange rates, firms use financial hedging
(Auboin & Ruta, 2012). In the view of the effect of exchange rate volatility on international
trade, it is essential to forecast the exchange rate in advance. This forms the core motivation of
our study.
2.2 Trade and Growth Theory
Trade implications of new growth theories are that trade and trade policies can influence
the long-run growth rate of the country. Trade and growth are interrelated to each other. There
9
Anonymous Manuscript
are three open economy versions of the canonical growth models, i.e., neoclassical growth
theory, learning-by-doing theory, and endogenous growth theory. Neoclassical growth model
focuses on factor accumulation. Learning-by-doing models emphasize accidental technological
progress. On the other hand, endogenous growth models center around motivated technological
progress. Trade enables the countries to import critical resources such as capital and intermediate
goods which have impact on long-run economic growth (Jayme Jr, 2001). These critical
resources are often quite expensive to produce locally. Authors have argued that openness to
trade, factor and technology flows, potentially contribute to the growth (Srinivasan & Bhagwati,
2001). Literature suggests that the exchange rate affects economic growth. Theory predicts that a
depreciating exchange rate boosts the exports of a country which will increase the gross domestic
product (GDP), but this holds mostly for developing countries (Rodrik, 2007). On the other hand,
developed countries are less likely to see benefits through the undervaluation of the exchange
rates. As the exchange rate has the significant impact of the economic growth of any country, it
is pertinent to study the forecasting of the exchange rates.
2.3 Financial Time Series Forecasting
The literature regarding forecasting financial time series is very vast. Hence, we would
present a brief review of the relevant literature which uses hybrid models. For linear data, ANNs
doesn’t perform better than ARIMA or random walk models. Because the data is having less
noise (disturbance), and we cannot expect ANNs to perform better than linear models for linear
relationships (Zhang et al., 1998). However, ARIMA and ANNs have gained success in their
linear and nonlinear domains, respectively. But none of them can be applied blindly for all
circumstances. It has become a standard practice to combine different models to overcome any
limitations posed by the component models and to improve the overall accuracy of the combined
10
Anonymous Manuscript
model. As it is very difficult to identify the linear and nonlinear characteristics of the data in the
real world, the combination of linear and nonlinear models could be a good strategy for practical
use (Khashei & Bijari, 2011). However, these hybrid models could not perform well in all the
circumstances (Taskaya-Temizel & Casey, 2005). In simple hybrid models, the first step is to
identify an ARIMA model to capture the linear structure of the data. In the second step, ANN is
trained on the residuals of the ARIMA model. Thus the future values are predicted by combining
both ANN and ARIMA.
Hybrid models could be homogeneous or heterogeneous. For example, if the hybrid model
is a combination of differently configured neural networks, it is termed as homogeneous.
Whereas, if both linear and nonlinear models are used to make a combination, it would be called
a heterogeneous hybrid model (Taskaya-Temizel & Casey, 2005). Both empirical, as well as
theoretical literature in the neural network forecasting, establish that the combining of different
methods or models could be an efficient way to improve the accuracy of the forecast
(Makridakis, 1989; Palm & Zellner, 1992). Various studies have used such hybrid models for
forecasting a time series (Luxhøj, Riis, & Stensballe, 1996; Wedding & Cios, 1996). In recent
years, several studies have used hybrid models which combine ARIMA and ANNs for modeling
and forecasting time series. Zhang (2003) uses hybrid models of ARIMA and ANNs on three
standard data sets, i.e., Wolf’s sunspot data, Canadian lynx data, and British pound/ US dollar
exchange rate data. The results reveal that the forecasting accuracy of hybrid models is better as
compared to that of either of the component models used separately. Khashei and Bijari (2010,
2011) propose a novel hybrid ARIMA-ANNs model, and the improved forecasting results are
obtained by such models for the same data sets used by Zhang (2003). Pai and Lin (2005)
combine ARIMA and support vector machines in forecasting stock prices problems, and finds
11
Anonymous Manuscript
promising results. Yu, Wang, and Lai (2005) propose a novel nonlinear ensemble forecasting
model integrating generalized linear auto-regression with ANNs to improve forecasting
performances in the foreign exchange market. Khashei, Reza Hejazi, and Bijari (2008) combine
ANNs with fuzzy regressions to forecast the financial market data, and the results show that such
models could be an effective alternate for forecasting a financial time series. Khashei, Bijari, and
Raissi Ardali (2009) integrate ARIMA models with ANNs and Fuzzy logic, and the findings
exhibit improved forecasting accuracy. Adhikari and Agrawal (2014) combine random walk
models with ANNs to four real-world financial time series, and the obtained results clearly show
that the hybrid model achieves better forecasting accuracies than individual component models.
Overall, it could be observed that the literature regarding forecasting the foreign exchange
rate is abundant. However, the studies employing hybrid models are very limited. This is one of
the first studies, to the best of our knowledge, which uses an ensemble model to forecast the
foreign exchange rate. The results of the study would be useful for policymakers, regulators,
investors, speculators, and arbitrageurs.
3 Data
The sample data in the current study consist of daily closing prices of three currency pairs –
Euro/US Dollar (EUR/USD), British Pound/ US Dollar (GBP/USD), and Japanese Yen/ US
Dollar (JPY/USD). The data has been obtained from the official website of the central banking
system of the United States which are certified by the Federal Reserve Bank of New York.1
1 Source: https://www.federalreserve.gov/releases/h10/hist/ [accessed on September 13, 2018]
12
Anonymous Manuscript
The sample period ranges from January 2009 to May 2016 (1848 observations). The data
before January 2009 is not considered deliberately to avoid the abnormal period of the subprime
crisis of 2008 which caused abnormal volatility in the market. Moreover, the inclusion of an
abnormal period could distort the results of the study. The returns for the currency pairs are
based on indirect exchange rate quotes of the foreign currencies (EUR, GBP, and JPY) with
respect to domestic currency (USD, and then logarithmic returns have been computed. The
logarithmic returns have been computed using equation (1) as follows:
Returnt=ln ( Pricet /Price t−1 )(1)
We divide the complete data into two subsets as the training and the testing datasets. The
period from January 2009 to March 2016 (1818 observations) is selected as the training set
which would be used to train the Neural Network model. The next 30 observations from the
months of April and May of 2016, which is not used in the training of the model, is set as the
testing set that would be used for checking the accuracy of the forecasting model.
The time series plots of all the three currency pairs are shown in Figure 1. The left side of
Figure 1 plots the prices of currencies at the level and the right side of the Figure plots the
logarithmic returns of that currency. It could be observed from Figure 1 that currency prices are
non-stationary, but their returns are stationary. To forecast a time-series, we need to ensure that
the time-series is stationary, i.e., the mean and the variance of the time-series are constant over
the sample period. We have employed Augmented Dickey-Fuller (ADF), and Phillips-Perron
(PP), unit root tests, to identify the stationarity properties of the exchange rates and returns. As
we are unaware of the true model for testing unit root, we would test two models: initially a
model with intercept only, and then a model with both intercept and trend. Optimal lag selection
for the ADF test is based upon Akaike information criterion (AIC). Currency exchange rates
13
Anonymous Manuscript
have been transformed into logarithmic values before applying the tests. The stationarity test
results have been reported in Table 1. We can reject the null hypothesis of non-stationarity, even
at 1% level of significance, for all three exchange rates after first differencing. Hence it could be
inferred that all three currency prices are non-stationary, but their returns are stationary.
<<Please insert Table 1 here>>
<<Please insert Figure 1 here>>
The descriptive statistics of the exchange rate returns are shown in Table 2. The mean daily
return is the lowest for EUR/USD, -0.011%. However, the volatility (standard deviation) is the
highest, 0.649%, for EUR/USD, which is very contradictory. Generally, the average mean has a
positive relationship with the volatility which can be observed in the other two currencies. The
mean daily returns for GBP/USD and JPY/USD are 0.000% and 0.010%, respectively with the
volatility of 0.599% and 0.647%, respectively. The skewness is negative for GBP/USD. Besides,
the kurtosis for GBP/USD pair is the highest (8.0951) among three pairs which suggest that the
GBP/USD pair has fat tails. The Jarque-Bera test statistic indicates that all three return series are
not distributed normally.
<<Please insert Table 2 here>>
4 Framework and Methodology
The study proposes a NAM Ensemble Framework as shown in Figure 2, for forecasting the
time series data of currency pairs. NAM stands for Neural Network, ARIMA, and Mean
Forecast. NAM Ensemble Framework consists of four phases- data collection, data
transformation, application of analytical models on the data, and the results phase. The data
collection phase employs the collection of the three most widely traded currency pairs-
14
Anonymous Manuscript
EUR/USD, GBP/USD, and JPY/USD. The data transformation employs distributing the data
sample into training and testing data. The data transformation also employs the computation of
returns for each pair. After that, the analytical models of Neural Network, ARIMA, and Mean
Forecast are applied to the transformed data.
<<Please insert Figure 2 here>>
Ensemble model has been constructed by integrating the three analytical forecasting models
using linear optimization approach. ARIMA and Mean Forecast capture the linearity associated
with the data, and the techniques rely on the assumption that the future values are considered to
be the linear function of the past observations. But the financial time series data consist of both
linear and nonlinear patterns. Therefore, the Neural Network captures the non-linearity
associated with the data. In this section, the basic concepts of the three components of the
ensemble model –Mean Forecast, ARIMA, and ANN along with ensemble model are briefly
reviewed. We have used three different models and a hybrid ensemble model for predicting the
exchange rate returns for three currency exchange pairs, i.e., EUR/USD, GBP/USD and
JPY/USD.
4.1 Mean Forecast Model
The mean forecast model gives the forecast value as the sample mean of the historical data.
The model returns forecasts for an independent and identically distributed (iid) model which is
shown by equation (2) as given below:
Y t=µ+Z t (2)
Where Zt is a normal iid error. The forecasted value from the model is given by equation (3) as
shown below:
15
Anonymous Manuscript
Y n (h )=µ(3)
Where µ is estimated by the sample mean. Thus, the mean forecast model is the very basic
model forecasting the sample mean as the future values. The mean forecast model is used in
forecasting of time series data and can sometimes prove to be the best model for forecasting.
4.2 Autoregressive Integrated Moving Average (ARIMA) Model
The ARIMA models are most popular in modeling a time series due to its ease of
implementation. In an ARIMA (p, d, q) model, the future values are considered to be the linear
function of several past observations, i.e., autoregressive (AR) terms and random shocks, i.e.,
moving average (MA) terms (Box & Jenkins, 1976). In ARIMA (p, d, q), p stands for AR terms,
d stands for the number of differences required to make a series stationary, and q stands for MA
terms. ARIMA models could be pure AR models, pure MA models, the combination of the AR
and the MA, i.e., ARMA models. The ARIMA models could be used for stationary time series.
The stationary time series implies that the mean and variance is constant over time. The ARIMA
methodology tries to identify the true data-generating process based on the time series
observations. A step-by-step procedure for ARIMA modeling is provided in Box and Jenkins
(1976). If a time series is stationary, the number of the AR and MA terms could be identified by
correlograms (autocorrelation and partial autocorrelation functions). The ARIMA model is
shown in equation (4) and has the following form:
Y t=α +∑i=1
p
β iY t−i+∑j=1
q
γ j εt− j+εt (4 )
16
Anonymous Manuscript
Where α is the intercept term, Y is a time series, t is the point of time, and ε t is a white noise.
These ε t are independent and identically distributed (i.i.d.) normal variables, which has zero
mean and a constant variance. P and q are the number of AR and MA terms, respectively.
The Box and Jenkins (1976) methodology is a hit and error technique with three iterative
steps. The first step is model identification, the second is parameter estimation, and the third step
is diagnostic checking. The ARIMA model is identified from autocorrelation and partial
autocorrelation properties of the time series. Then, possible models are estimated, and the best
model is selected by some information criterion (such as Akaike information criterion). Before
identifying the model, it is necessary that the series under investigation is stationary. After
estimating the model, the adequacy of the model is tested by diagnostic tests. These tests are
basically to check whether the residuals of the model are i.i.d. This entire process is repeated
until a suitable model, satisfying the assumption about random errors, is found. The final
selected model can be used for the forecasting purpose.
4.3 Neural Network (NN) Model
Neural Network (NN) has the structure of the biological neural network and the learning
capabilities which can help in predicting future values based on past data. NN, as a machine
learning technique, is used in varied domains like credit ratings and approvals (Huang et al.,
2004), stock market prediction (Grudnitski & Osburn, 1993), debt risk assessment, and
bankruptcy prediction (Atiya, 2001; Kumar & Ravi, 2007; Wilson & Sharda, 1994). NN models
find its applications in both economic and financial time-series (Bildirici, Alp, & Ersin, 2010;
Bildirici & Ersin, 2009; Dunis et al., 2012; Özkan, 2013; Zhu, Wang, Xu, & Li, 2008).
17
Anonymous Manuscript
The NN model has an advantage over other models because it can capture the non-linearity
associated with the time series data of exchange rate returns. NN models have the function
mapping capability and are also referred to as universal approximators (Chen, Leung, & Daouk,
2003; Zhang & Qi, 2005). NN model can approximate a large number of inputs. The
configuration of neural network model consists of an input layer, hidden layers, and output layer
and the forecasting accuracy of NN model depends on both, number of inputs as well as the
configuration of NN model. The number of inputs determines the lag in the NN model. We have
used the feed-forward neural network in this study. The NN model used a back-propagation
learning method and is initially trained for forecasting only one-step ahead exchange rate. Multi-
step forecasts for exchange rate are then computed iteratively. The NN model is shown in
equation (5) and has the following form:
Y t=α 0+∑j=1
q
α j g (β0 j+∑i=1
p
β ij y t−i)+εt(5)
Where α j (j=0,1,2…,q) and β ij (i=0,1,2,…,p; j=1,2,..,q) are model parameters known as
connection weights; p and q are number of nodes in input and hidden layer; g is the transfer
function, which can be represented by equation (6).
g ( x )= 11+e−x (6)
The neural network model uses the exchange rates of a specific currency pair (EUR/USD,
GBP/USD and JPY/USD) as input nodes. The number of optimum lags is obtained as per the
AIC. The lag length for EUR/USD according to AIC comes out to be 2. Similarly, the lag length
of JPY/USD comes out to be 3, and for GBP/USD, the lag length comes out to be only 4. The lag
length in time series forecasting implies that a total of n nodes (equal to the lag length) would be
18
Anonymous Manuscript
used as inputs in the input layer of the feed-forward neural network. For instance, EUR/USD
based neural network has a lag length of 2 units, which implies that total two units would be used
as input to the neural network. Thus, the configuration of the neural network for EUR/USD
consists of 2-1-1 configuration. Hence, there would be overall three neural network models for
EUR/USD, GBP/USD, and JPY/USD currency pairs having the configurations of 2-1-1, 4-2-1
and 3-2-1 respectively. The generic neural network configuration of NN can be seen in Figure 3.
<<Please insert Figure 3 here>>
4.4 Ensemble Model
The ensemble model or hybrid model comprises of combining individual models to improve
the forecasting (Van Wezel & Potharst, 2007). Ensemble model produces better results than the
individual models whenever there is diversity in the dataset and the models (Coussement & De
Bock, 2013). Ensemble model has gained importance because of its wider applicability in
various domains (Bauer & Kohavi, 1999). The ensemble model gained its popularity in varied
domains primarily because of its simplicity to combine model with favourable performance and
yield better results. Ensemble model has been applied in both cross-sectional as well as time
series data. Similarly, the ensemble model has also been used in time series data for forecasting
purposes (Adhikari & Agrawal, 2014; Khashei & Bijari, 2011; Zhang, 2003).
Prediction of ensemble model in financial domain can be mainly grouped into two
categories: (i) segregating the linear and non-linear part in the time series data and then using one
forecasting model to capture the linear part and using other forecasting model to capture the
nonlinear part, then finally combining the two forecast model to yield the actual forecasted value
(Adhikari & Agrawal, 2014; Khashei & Bijari, 2011; Zhang, 2003); and (ii) the forecasted values
from each model are given weights to yield the final forecast value (Shmueli & Koppius, 2010).
19
Anonymous Manuscript
The weights of each individual model could be given based on different methods like linear
optimization, analytical hierarchical processing, fuzzy sets, etc.
The current study employs the second type of ensemble models, where weights are obtained
by combining individual models through linear optimization. We have constructed an ensemble
of three models: ARIMA, Neural Network, and Mean Forecast. The weights (w1,w2, and w3) for
each of the three ensemble model (EUR/USD, GBP/USD and JPY/USD) were obtained by linear
optimization approach. The objective function in linear optimization approach is to minimize the
three error terms (RMSE, MAPE, and MAE) separately, keeping the constraint as the sum of
weights for each model should be unity. After that, the weights thus obtained are averaged across
the error terms to yield the final weight of each model in the ensemble.
5 Results
For comparing the efficiency of the neural network, we consider three widely used error
metrics – Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute
Percentage Error (MAPE) – as shown in Equation (7), (8) and (9) respectively.
RMSE=√∑i=1
n
(Y i−Y i )2
n (7)
MAE=∑i=1
n
|Y i−Y i|
n (8)
(9)
20
Anonymous Manuscript
Where Y is the value of the time-series of the exchange rate returns. The lesser the value of
RMSE, MAE, and MAPE, the better is the accuracy of the model in forecasting the exchange
rate of the currency.
5.1 Ensemble for EUR/USD Currency Exchange Rate
<<Please insert Figure 4 here>>
The ensemble model for EUR/USD has a weightage of 33.33% for neural network model
and 66.67% for ARIMA model whereas Mean Forecast model has zero weightage in the
ensemble as shown in Figure 4. This implies that ARIMA has a higher weightage in the
composition of the ensemble model. The results for the EUR/USD currency exchange rate by
their error terms (RMSE, MAE, and MAPE) can be seen in Table 3.
<<Please insert Table 3 here>>
We have taken six different horizons for the forecast (h=5, 10, 15, 20, 25 and 30). The last row
of each subpart in the table gives the average value of all forecasted horizons and ranking of a
particular model. The bold-faced values in each table indicate the best performing model. Thus,
an ensemble model with average RMSE value equal to 0.003750 seems to be the best performing
model for RMSE error metrics. The ensemble model closely competes with the neural network
model with RMSE value of 0.003753. Therefore, we can say that both neural network, as well as
the ensemble model, are able to forecast better for EUR/USD exchange rate in terms of RMSE.
For MAE, the bold-faced value of 0.00256 suggests that the Mean Forecast model is able to
predict better than all the other models. Mean Forecast model yield the average value of MAE as
0.00256, which is way better than the next competing models of ARIMA and Ensemble, having
21
Anonymous Manuscript
the MAE of 0.00278. Thus, the Mean Forecast model is 8.59% better than its competing models
of ARIMA and Ensemble.
The results from Table 3 indicates that the ARIMA model is able to predict better than all
other models in terms of MAPE. The ARIMA model is considered better than other models
because the model yields the minimum value of MAPE, i.e., 104.6022, across all models. The
ARIMA model closely competes with the ensemble model, with MAPE value of 105.2602.
Therefore, we can say that both ARIMA, as well as Ensemble model, are able to forecast better
for EUR/USD exchange rate in terms of MAPE.
The models are ranked based on their error metrics. Thus, the ensemble model emerges as
the best model in terms of RMSE. Whereas the mean forecast model emerges as the best model
in terms of MAE, and ARIMA model emerges as the best model for MAPE. The overall average
rank obtained in Table 4 indicates that the ensemble model with an average ranking of 1.67
emerges as the best model for forecasting the returns of EUR/USD currency exchange rates.
5.2 Ensemble for GBP/USD Currency Exchange Rate
<<Please insert Figure 5 here>>
The ensemble model for GBP/USD has a weightage of just 11.33% for neural network and
88.67% for Mean Forecast model whereas ARIMA model does not have any significance in the
construction of ensemble model as shown in Figure 5.
The results of Table 4 indicates that the ensemble model is able to predict better than all
other models in terms of all the three error metrics, i.e., RMSE, MAE, and MAPE. The ensemble
model is considered best among all the models because the ensemble model yields the minimum
value of RMSE, MAE as well as MAPE. The ensemble model has the value of RMSE as
22
Anonymous Manuscript
0.00618, MAE as 0.00508 and MAPE as 97.99. Therefore, we can say that the ensemble model
is able to forecast better for GBP/USD exchange rate.
<<Please insert Table 4 here>>
Thus, the overall average rank obtained in Table 4 indicates that the ensemble model with an
average ranking of 1 emerges as the best model for forecasting the returns of the GBP/USD
currency exchange rates. This suggests that the performance of the ensemble model is more
robust and better than the individual models.
5.3 Ensemble for JPY/USD Currency Exchange Rate
<<Please insert Figure 6 here>>
The ensemble model for JPY/USD has a full weightage of 100% for neural network whereas
ARIMA and Mean Forecast model does not have any weightage in the ensemble as shown in
Figure 6. This can be attributed to the volatility of the exchange rate returns in the testing period.
The results of Table 5 indicates that the neural network model, the mean forecast model, and
the ensemble model are able to predict the JPY/USD currency exchange rate. The main reason
for having the same values of error metrics for both neural network model and the ensemble
model is that the neural network model has full 100% weightage in the ensemble model. The
neural network model and ensemble model yield the value of RMSE, MAE, and MAPE as
0.00801, 0.00619 and 104.6062 respectively.
<<Please insert Table 5 here>>
The models are ranked based on their error metrics. Thus, the ensemble model along with
Neural network model and Mean Forecast model emerges as the best model, and the
23
Anonymous Manuscript
performance of the ARIMA model is less than the performance of other three models. The
overall average rank obtained in Table 5 indicates that ensemble model, NN model and hybrid
model with an average ranking of 1.67 emerge as the best models for forecasting the returns of
JPY/USD currency exchange rates.
6 Discussion and Policy Implications
This study proposes a novel ensemble approach for exchange rate forecast. Ensemble
technique incorporates both linear and non-linear methods for exchange rate forecast. It assigns
weights to different linear and non-linear methods according to their prediction power. Ensemble
model comprises of different models to improve the forecasting power. We have used an
ensemble technique for exchange rate prediction for the three most traded currency pairs
(EUR/USD, GBP/USD, and JPY/USD). As it is evident from the hypothetical case, the exchange
rate forecasting is essential for the evaluation of risk and return associated with international
trade.
This study highlights that the performance of the ensemble model is more robust and better
than the individual models for all the three currency pairs, i.e., EUR/USD, GBP/USD, and
JPY/USD. In this study, we have formed an ensemble model using three different models, i.e.,
NN, ARIMA, and Mean Forecast model. Ensemble model gives a different set of weights to
these models for different currency pairs depending upon the predictive power of these models
for each currency pair. Moreover, the weight of NN in the ensemble model is highest (100%) for
the JPY-USD currency pair because the volatility of the JPY-USD pair is highest for the same
period. This confirms the fact that predictive power of NN is good for the volatile currency pair
(Inani, Tripathi, & Kumar, 2016). We have taken currency pairs of world’s largest economies
which will help investors and governments to evaluate the risk and return relationship of their
24
Anonymous Manuscript
business decisions. Exchange rate forecasting is crucial in today’s economic scenario because
exchange rate volatility has a direct impact on international trade which in turn affects the world
economy.
A hypothetical case for the policy makers
Let us assume that an international trader has made a transaction worth USD 1 billion. Now,
we would show the economic benefits earned by employing the ensemble forecasting technique
vis-à-vis worst performing predictive model. The difference between the actual and predicted
price of the transaction would depend on the forecasting method we are using. We have
considered here the 30-days horizon (of the testing period of this study) for computing the
economic edge offered by employing an ensemble method. The forecasted prices from
forecasted returns have been computed by equation (10) derived from equation (1) as shown
below:
Returnt=ln ( Pricet /Price t−1 )(1)
Hence, Pricet / Pricet−1=antilog (Return¿¿ t )=eReturn t¿
Pricet=eR eturn t× Pricet−1(10)
Now, for computing the price at a particular time (t), we need to have both returns at the
time (t) and the price at the time (t-1). The actual price of EUR/USD on March 31, 2016, is 1.139
and the returns computed by the ARIMA model and ensemble model for April 1, 2016, are -
0.00032 and -0.00025 respectively. Thus, the price of EUR/USD on April 1, 2016, would be
1.13863 (=1.139*e−0.00032) for ARIMA model and 1.13871 (=1.139*e−0.00025). Likewise, we have
calculated the price for next 30-days. The details of the same can be seen in appendix 1.
25
Anonymous Manuscript
After computing the forecasted prices, the MAE is computed. The computation of economic
benefit offered by the ensemble model is presented below step-wise for the EUR/USD currency
pair for 30-days testing period.
The value of transaction = USD 1 billion (as mentioned above)
Predicted value of EUR/USD by ARIMA = 0.005897158 euro/dollar
Predicted value of EUR/USD by Ensemble = 0.00584471 euro/dollar
Difference in predicted value by ensemble and ARIMA for USD 1 billion
= (0.005897158 - 0.00584471)* USD 1 billion = 52,448 euro = 0.052 Million euro
Similarly, we have calculated the economic benefit for other currency-pairs as shown in Table 6.
<<Please insert Table 6 here>>
The economic edge of the ensemble is calculated by computing the difference between
predicted prices of the ensemble with respect to the worst performing predictive model. The
results of Table 6 suggest that the ensemble model improves the performance of the base model
by giving an economic edge of 0.052 million EUR, 3.13 million GBP, and 76.5 million JPY per
billion dollar transactions respectively. Along with traders, the ensemble model might give the
economic edge to arbitrageurs and speculators indulged in the foreign exchange market.
7 Economic Policy Implications
The behavior of the currency exchange market (FOREX) is irregular and random. The
reason behind that is exchange rates are dependent on many economic and political factors.
Hence, exchange rate prediction is not a simple task. The next obvious question arises “what is
the need to forecast the exchange rate?” This question has drawn the attention of researchers and
26
Anonymous Manuscript
policymakers for many years. Investors and arbitrageurs follow exchange rate fluctuations and
get benefits from currency trading. Exchange rate prediction enables the policy makers to take an
appropriate decision about future investments. With the advent of floating exchange rate and
abandonment of the Bretton Woods System, the exchange rate is the principal channel through
which monetary policy affects the inflation and other economic activities. Floating exchange rate
mechanism is aimed to empower the policy makers to form the monetary policy independent of
external economic inequities. A better understanding of exchange rate movements is essential to
curb the inflation at a moderate level and keep the economic activities at a higher level.
Exchange rate reflects the financial and economic condition of the country. Hence,
understanding of the movement of the exchange rate will enable the policy makers to get timely
information about the financial and economic condition of the economy of a country.
Exchange rate prediction is an essential constituent of monetary policy of the national
economies especially in the context of leading market economies. These economies are very
sensitive to external negative news. Recently, Eurozone debt crisis, change in fuel prices, global
financial crisis, and tapering of US Federal Reserve’s quantitative easing have affected the
economic health of leading economies in the world. These negative news result in uncertainty in
economic world which in turn result in high volatility in the currencies of these countries. This
uncertainty in the movements of currency has led to the foundation of exchange rate prediction.
Exchange rate forecasting becomes important in the context of high-value economic transactions
in the form of debt and equity from multinational corporations and portfolio investors. This study
helps all the stakeholders such as central bank, government, and investors who are affected by
exchange rate volatility. The Central bank of a country intervenes in the forex market in case of
disruptive or undesirable movements in the exchange rates which may be unhealthy for the
27
Anonymous Manuscript
internal and external sector of the economy (Kayal & Maheswaran, 2016; Prakash, 2012).
Central bank takes some quantitative as well as qualitative measures to ease down the volatility
of the currency. Hence, exchange rate prediction is essential for the central bank to bring the
desired economic stability in the country by ensuring currency behaving steadily without any
sharp moves.
The ensemble model incorporates both the linear as well as the nonlinear behavior of
exchange rate movements which may help the policy makers and economist to form a suitable
monetary policy for price stability and better economic activity. The finding further suggests that
the ensemble model better extracts the information hidden in the exchange rate which in turn
help the monetary authorities to have a better understanding of the financial and economic
condition of the economy. Better exchange rate prediction will help investors to form a profitable
trading strategy for currency trading.
8 Academic Implications
Researchers have been facing major challenges in devising an effective model for exchange
rate forecasting. Exchange rates often show irregular patterns which are difficult to predict for
researchers. This study proposes a new ensemble model which intelligently combines the three
models NN, ARIMA, and mean forecast model to forecast exchange rate. Empirical results
suggest that the ensemble model substantially improves the overall forecasting accuracies and
also outperforms each of the component models separately. This study contributes to the
literature of exchange rates prediction based on hybrid model. Earlier studies on hybrid model
mostly take equal weights for all the component models, but this study assigns the weights
according to the predictive power of the different component models present in the ensemble.
This finding can be an important implication for future researchers who want to predict the
28
Anonymous Manuscript
exchange rate using some ensemble model. Moreover, this study highlights the fact that the
weight of NN in the ensemble model is highest for the most volatile currency pair, here
JPY/USD.
9 Conclusions, Limitations, and Future Research
The accuracy of exchange rate forecasting has been a crucial yet challenging task for
researchers in the past. Authors have used various time series models for exchange rate
prediction, yet research for effectiveness of forecasting models is still going. It has been
observed that combining predictive power from different models often leads to improved
performance. Hence, in this study, we have used an ensemble model comprising of three
different models, i.e., NN, ARIMA, and mean forecast model. In this study, we aim to forecast
the exchange rate for three currency pairs, i.e., EUR/USD, GBP/USD, and JPY/USD for the
period from January 2009 to May 2016. The results ascertain that the predictive power of
ensemble model is best for all the three currency pairs as compared to other models such as NN,
ARIMA, and mean forecast model. Moreover, when the exchange rate volatility is high, weight
for NN is coming to be highest in the ensemble model. This study contributes to the literature on
exchange rate prediction. Exchange rate prediction is imperative for various stakeholders such as
government, the central bank, and arbitrageurs to take their decision with minimum risk and
maximum returns.
We have taken only three currency pairs for exchange rate prediction. Future researchers can
try other combinations to examine the predictive power of the ensemble model. This will help
investors and government to take an appropriate financial decision in advance. We have used
ARIMA, NN and mean forecast models for this study; future researchers can use other
combination models in the ensemble. They can use other variants of NN such as Recurring NN
29
Anonymous Manuscript
(RNN) and Higher Order NN (HONN) in the ensemble model. We have used data for the period
from January 2009 to May 2016 in our study. Future researchers may take the data before
January 2009 to observe the prediction power of the ensemble model in the period of the
subprime crisis of 2008 which caused abnormal volatility in the market.
Appendix 1
Calculation of price differences for EUR/USD (30-days testing period)
Date Actual Price
Returns(ARIMA)
Returns(ensemble)
Price(ARIMA)
Price(ensemble)
01-Apr-16 1.1385 -0.00032 -0.0002507 1.138637 1.13871504-Apr-16 1.1386 -9.6E-05 -9.216E-05 1.138528 1.1386105-Apr-16 1.1374 -5.9E-05 -6.786E-05 1.13846 1.13853206-Apr-16 1.143 -0.00014 -0.0001247 1.138295 1.1383907-Apr-16 1.1386 -0.00011 -0.0001009 1.138171 1.13827508-Apr-16 1.1406 -0.0001 -9.719E-05 1.138053 1.13816511-Apr-16 1.1412 -0.00012 -0.0001062 1.13792 1.13804412-Apr-16 1.1395 -0.00011 -0.0001024 1.137794 1.13792813-Apr-16 1.1281 -0.00011 -0.0001018 1.137668 1.13781214-Apr-16 1.1262 -0.00011 -0.0001032 1.13754 1.13769415-Apr-16 1.1295 -0.00011 -0.0001026 1.137413 1.13757718-Apr-16 1.1322 -0.00011 -0.0001025 1.137286 1.13746119-Apr-16 1.1375 -0.00011 -0.0001028 1.137159 1.13734420-Apr-16 1.133 -0.00011 -0.0001027 1.137032 1.13722721-Apr-16 1.1301 -0.00011 -0.0001027 1.136905 1.1371122-Apr-16 1.1239 -0.00011 -0.0001027 1.136778 1.13699425-Apr-16 1.1274 -0.00011 -0.0001027 1.136651 1.13687726-Apr-16 1.1318 -0.00011 -0.0001027 1.136524 1.1367627-Apr-16 1.1322 -0.00011 -0.0001027 1.136398 1.13664328-Apr-16 1.1325 -0.00011 -0.0001027 1.136271 1.13652729-Apr-16 1.1441 -0.00011 -0.0001027 1.136144 1.1364102-May-16 1.1516 -0.00011 -0.0001027 1.136017 1.13629303-May-16 1.1508 -0.00011 -0.0001027 1.13589 1.13617704-May-16 1.1486 -0.00011 -0.0001027 1.135763 1.1360605-May-16 1.1404 -0.00011 -0.0001027 1.135636 1.13594306-May-16 1.1421 -0.00011 -0.0001027 1.135509 1.13582709-May-16 1.1402 -0.00011 -0.0001027 1.135383 1.13571
30
Anonymous Manuscript
10-May-16 1.1386 -0.00011 -0.0001027 1.135256 1.13559411-May-16 1.1444 -0.00011 -0.0001027 1.135129 1.13547712-May-16 1.138 -0.00011 -0.0001027 1.135002 1.13536
References
Adhikari, R., & Agrawal, R. K. (2014). A combination of artificial neural network and random
walk models for financial time series forecasting. Neural Computing and Applications,
24(6), 1441–1449.
Alfaro, E., García, N., Gámez, M., & Elizondo, D. (2008). Bankruptcy forecasting: An empirical
comparison of AdaBoost and neural networks. Decision Support Systems, 45(1), 110–
122.
Atiya, A. F. (2001). Bankruptcy prediction for credit risk using neural networks: A survey and
new results. IEEE Transactions on Neural Networks, 12(4), 929–935.
Auboin, M., & Ruta, M. (2012). The Relationship between Exchange Rates and International
Trade: A Literature Review (CESifo Working Paper Series No. 3868). CESifo Group
Munich.
31
Anonymous Manuscript
Balkin, S. D., & Ord, J. K. (2000). Automatic neural network modeling for univariate time
series. International Journal of Forecasting, 16(4), 509–515.
Baron, D. P. (1976). Flexible exchange rates, forward markets, and the level of trade. The
American Economic Review, 66(3), 253–266.
Bauer, E., & Kohavi, R. (1999). An Empirical Comparison of Voting Classification Algorithms:
Bagging, Boosting, and Variants. Machine Learning, 36(1–2), 105–139.
Bildirici, M., Alp, E. A., & Ersin, Ö. Ö. (2010). TAR-cointegration neural network model: An
empirical analysis of exchange rates and stock returns. Expert Systems with Applications,
37(1), 2–11.
Bildirici, M., & Ersin, Ö. Ö. (2009). Improving forecasts of GARCH family models with the
artificial neural networks: An application to the daily returns in Istanbul Stock Exchange.
Expert Systems with Applications, 36(4), 7355–7362.
Box, G. E., & Jenkins, G. M. (1976). Time series analysis, control, and forecasting. Holden-Day
Inc., San Francisco, CA.
Broll, U., & Eckwert, B. (1999). Exchange rate volatility and international trade. Southern
Economic Journal, 178–185.
Buckley, P., & O’Brien, F. (2017). The effect of malicious manipulations on prediction market
accuracy. Information Systems Frontiers, 19(3), 611–623.
Byrne, J. P., Darby, J., & MacDonald, R. (2008). US trade and exchange rate volatility: A real
sectoral bilateral analysis. Journal of Macroeconomics, 30(1), 238–259.
Chen, A.-S., Leung, M. T., & Daouk, H. (2003). Application of neural networks to an emerging
financial market: forecasting and trading the Taiwan Stock Index. Computers &
Operations Research, 30(6), 901–923.
32
Anonymous Manuscript
Choi, T.-M., Yu, Y., & Au, K.-F. (2011). A hybrid SARIMA wavelet transform method for sales
forecasting. Decision Support Systems, 51(1), 130–140.
Clark, P. B. (1973). Uncertainty, exchange risk, and the level of international trade. Economic
Inquiry, 11(3), 302–313.
Coussement, K., & De Bock, K. W. (2013). Customer churn prediction in the online gambling
industry: The beneficial effect of ensemble learning. Journal of Business Research, 66(9),
1629–1636.
Cushman, D. O. (1983). The effects of real exchange rate risk on international trade. Journal of
International Economics, 15(1–2), 45–63.
[Dataset] Federal Reserve Bank, 2018, Foreign Exchange Rates - H.10,
https://www.federalreserve.gov/releases/h10/hist/ [accessed on September 13, 2018]
Dornbusch, R., Fischer, S., & Samuelson, P. A. (1977). Comparative advantage, trade, and
payments in a Ricardian model with a continuum of goods. The American Economic
Review, 67(5), 823–839.
Dunis, C. L., Laws, J., & Schilling, U. (2012). Currency trading in volatile markets: Did neural
networks outperform for the EUR/USD during the financial crisis 2007–2009? Journal of
Derivatives & Hedge Funds, 18(1), 2–41.
Ghazali, R., Jaafar Hussain, A., Mohd Nawi, N., & Mohamad, B. (2009). Non-stationary and
stationary prediction of financial time series using dynamic ridge polynomial neural
network. Neurocomputing, 72(10–12), 2359–2367.
Grudnitski, G., & Osburn, L. (1993). Forecasting S&P and gold futures prices: An application of
neural networks. Journal of Futures Markets, 13(6), 631–643.
33
Anonymous Manuscript
Hann, T. H., & Steurer, E. (1996). Much ado about nothing? Exchange rate forecasting: Neural
networks vs. linear models using monthly and weekly data. Neurocomputing, 10(4), 323–
339.
Heckscher, E. F., & Ohlin, B. G. (1991). Heckscher-Ohlin trade theory. The MIT Press.
Huang, Z., Chen, H., Hsu, C.-J., Chen, W.-H., & Wu, S. (2004). Credit rating analysis with
support vector machines and neural networks: a market comparative study. Decision
Support Systems, 37(4), 543–558.
Hussain, A. J., Knowles, A., Lisboa, P. J. G., & El-Deredy, W. (2008). Financial time series
prediction using polynomial pipelined neural networks. Expert Systems with
Applications, 35(3), 1186–1199.
Inani, S. K., Tripathi, M., & Kumar, S. (2016). Does Artificial Neural Network Forecast Better
for Excessively Volatile Currency Pairs? Journal of Prediction Markets, 10(2).
Ince, H., & Trafalis, T. B. (2006). A hybrid model for exchange rate prediction. Decision
Support Systems, 42(2), 1054–1062.
Jayme Jr, F. G. (2001). Notes on trade and growth. Texto Para Discussão, (166).
Kayal, P., & Maheswaran, S. (2016). Is USD-INR Really an Excessively Volatile Currency Pair?
Journal of Quantitative Economics, 1–14.
Khansa, L., & Liginlal, D. (2011). Predicting stock market returns from malicious attacks: A
comparative analysis of vector autoregression and time-delayed neural networks.
Decision Support Systems, 51(4), 745–759.
Khashei, M., & Bijari, M. (2010). An artificial neural network (p, d, q) model for timeseries
forecasting. Expert Systems with Applications, 37(1), 479–489.
34
Anonymous Manuscript
Khashei, M., & Bijari, M. (2011). A novel hybridization of artificial neural networks and
ARIMA models for time series forecasting. Applied Soft Computing, 11(2), 2664–2675.
Khashei, M., Bijari, M., & Raissi Ardali, G. A. (2009). Improvement of Auto-Regressive
Integrated Moving Average models using Fuzzy logic and Artificial Neural Networks
(ANNs). Neurocomputing, 72(4–6), 956–967.
Khashei, M., Reza Hejazi, S., & Bijari, M. (2008). A new hybrid artificial neural networks and
fuzzy regression model for time series forecasting. Fuzzy Sets and Systems, 159(7), 769–
786.
Kumar, P. R., & Ravi, V. (2007). Bankruptcy prediction in banks and firms via statistical and
intelligent techniques – A review. European Journal of Operational Research, 180(1), 1–
28.
Lam, M. (2004). Neural network techniques for financial performance prediction: integrating
fundamental and technical analysis. Decision Support Systems, 37(4), 567–581.
Lee, J. K., & Yum, C. S. (1998). Judgemental adjustment in time series forecasting using neural
networks. Decision Support Systems, 22(2), 135.
Luxhøj, J. T., Riis, J. O., & Stensballe, B. (1996). A hybrid econometric—neural network
modeling approach for sales forecasting. International Journal of Production Economics,
43(2), 175–192.
Makridakis, S. (1989). Why combining works? International Journal of Forecasting, 5(4), 601–
603.
Özkan, F. (2013). Comparing the forecasting performance of neural network and purchasing
power parity: The case of Turkey. Economic Modelling, 31, 752–758.
35
Anonymous Manuscript
Pai, P.-F., & Lin, C.-S. (2005). A hybrid ARIMA and support vector machines model in stock
price forecasting. Omega, 33(6), 497–505.
Palm, F. C., & Zellner, A. (1992). To combine or not to combine? issues of combining forecasts.
Journal of Forecasting, 11(8), 687–701.
Panda, C., & Narasimhan, V. (2007). Forecasting exchange rate better with artificial neural
network. Journal of Policy Modeling, 29(2), 227-236.
Pasley, A., & Austin, J. (2004). Distribution forecasting of high frequency time series. Decision
Support Systems, 37(4), 501–513.
Péridy, N. (2003). Exchange rate volatility, sectoral trade, and the aggregation bias. Review of
World Economics, 139(3), 389–418.
Prakash, A. (2012). Major Episodes of Volatility in the Indian Foreign Exchange Market in the
Last Two Decades (1993-2013): Central Bank’s Response, 33(1 & 2), 166–199.
Prusa, J. D., Sagul, R. T., & Khoshgoftaar, T. M. (2018). Extracting Knowledge from Technical
Reports for the Valuation of West Texas Intermediate Crude Oil Futures. Information
Systems Frontiers.
Qureshi, S., Rehman, I. U., & Qureshi, F. (2018). Does gold act as a safe haven against exchange
rate fluctuations? The case of Pakistan rupee. Journal of Policy Modeling, 40(4), 685-
708.
Rodríguez-González, A., Colomo-Palacios, R., Guldris-Iglesias, F., Gómez-Berbís, J. M., &
García-Crespo, A. (2012). FAST: Fundamental Analysis Support for Financial
Statements. Using semantics for trading recommendations. Information Systems
Frontiers, 14(5), 999–1017.
36
Anonymous Manuscript
Rodrik, D. (2007). The Real Exchange Rate and Economic Growth: Theory and Evidence,”
Kennedy School of Government manuscript. In Center for Global Development.
Sermpinis, G., Dunis, C., Laws, J., & Stasinakis, C. (2012). Forecasting and trading the
EUR/USD exchange rate with stochastic Neural Network combination and time-varying
leverage. Decision Support Systems, 54(1), 316–329.
Shmueli, G., & Koppius, O. (2010). Predictive analytics in information systems research. Robert
H. Smith School Research Paper No. RHS, 06–138.
Srinivasan, T. N., & Bhagwati, J. (2001). Outward-orientation and development: are revisionists
right? In Trade, development and political economy (pp. 3–26). Springer.
Taskaya-Temizel, T., & Casey, M. C. (2005). A comparative study of autoregressive neural
network hybrids. Neural Networks, 18(5–6), 781–789.
Tkáč, M., & Verner, R. (2016). Artificial neural networks in business: Two decades of research.
Applied Soft Computing, 38, 788–804.
Van Wezel, M., & Potharst, R. (2007). Improved customer choice predictions using ensemble
methods. European Journal of Operational Research, 181(1), 436–452.
Wedding, D. K., & Cios, K. J. (1996). Time series forecasting by combining RBF networks,
certainty factors, and the Box-Jenkins model. Neurocomputing, 10(2), 149–168.
Wilson, R. L., & Sharda, R. (1994). Bankruptcy prediction using neural networks. Decision
Support Systems, 11(5), 545–557.
Yu, L., Wang, S., & Lai, K. K. (2005). A novel nonlinear ensemble forecasting model
incorporating GLAR and ANN for foreign exchange rates. Computers & Operations
Research, 32(10), 2523–2541.
37
Anonymous Manuscript
Zhang, G., Eddy Patuwo, B., & Y. Hu, M. (1998). Forecasting with artificial neural networks::
The state of the art. International Journal of Forecasting, 14(1), 35–62.
Zhang, G. P. (2003). Time series forecasting using a hybrid ARIMA and neural network model.
Neurocomputing, 50, 159–175.
Zhang, G. P., & Qi, M. (2005). Neural network forecasting for seasonal and trend time series.
European Journal of Operational Research, 160(2), 501–514.
Zhu, X., Wang, H., Xu, L., & Li, H. (2008). Predicting stock index increments by neural
networks: The role of trading volume under different horizons. Expert Systems with
Applications, 34(4), 3043–3054.
38
Figures
1.0
1.1
1.2
1.3
1.4
1.5
1.6
09 10 11 12 13 14 15 16
EUR/USD
-.04
-.02
.00
.02
.04
.06
09 10 11 12 13 14 15 16
Ret_EUR/USD
1.3
1.4
1.5
1.6
1.7
1.8
09 10 11 12 13 14 15 16
GBP/USD
-.06
-.04
-.02
.00
.02
.04
.06
09 10 11 12 13 14 15 16
Ret_GBP/USD
70
80
90
100
110
120
130
09 10 11 12 13 14 15 16
JPY/USD
-.06
-.04
-.02
.00
.02
.04
09 10 11 12 13 14 15 16
Ret_JPY/USD
Figure 1: Plots of prices and returns of the currencies. The left side of the Figure plots the
prices of currencies at the level, and the right side of the Figure plots the logarithmic
returns of that currency.
39
Figures
Figure 2: NAM Ensemble framework
Figure 3: Configuration of Neural Network model
40
Figures
Figure 4: Ensemble model for EUR/USD currency exchange rate
Figure 5: Ensemble model for the GBP/USD currency exchange rate
Figure 6: Ensemble model for JPY/USD currency exchange rate
41
Figures
42
Tables
Table 1. Stationarity test results
logarithm of EUR/USD logarithm of GBP/USD logarithm of JPY/USDModel and test Level FD Level FD Level FDWith intercept onlyADF -1.39 -43.29*** -2.91 -17.80*** -0.74 -31.72***PP -1.39 -43.31*** -2.83 -42.56*** -0.71 -43.02***With intercept and trendADF -2.26 -43.28*** -3.04 -17.83*** -1.57 -31.73***PP -2.23 -43.30*** -2.96 -42.58*** -1.55 -43.02***Notes: FD stands for first differences. *** indicates significance at 1% level of confidence. Only t-statistics
have been reported in this table to conserve space. The critical value for 1% level of significance is -3.43 for a
model with intercept only, whereas the critical value for a model with intercept and trend is -3.96 for 1%
level of significance.
Table 2. Descriptive statistics of daily exchange rate returns (January 2009 to May 2016)
EUR/USD GBP/USD JPY/USD Mean % -0.011 0.000 0.010 Median % 0.000 0.013 0.010 Maximum % 4.621 4.273 3.343 Minimum % -2.689 -4.966 -4.409 Standard Deviation % 0.649 0.599 0.647 Skewness 0.1842 -0.1652 0.0933 Kurtosis 5.3431 8.0951 6.7207
Jarque-Bera 432.9604 2006.213 1068.072 Probability 0.00 0.00 0.00 Observations 1847 1847 1847
Table 3. Model comparison for EUR/USD currency pairs
43
Tables
Error MetricsForecast Horizon
Neural Network ARIMA Mean
Forecast Ensemble
RMSE h=5 0.00284 0.00285 0.00383 0.00284h=10 0.00384 0.00384 0.00384 0.00384h=15 0.00371 0.00371 0.00383 0.00371h=20 0.00362 0.00362 0.00383 0.00362h=25 0.00431 0.00431 0.00383 0.00431h=30 0.0042 0.0042 0.00383 0.0042
Average 0.003753 0.00376 0.00383 0.00375Rank 2 3 4 1
MAE h=5 0.00204 0.00202 0.00256 0.00202h=10 0.00256 0.00256 0.00256 0.00255h=15 0.00281 0.0028 0.00256 0.00281h=20 0.00278 0.00277 0.00256 0.00278h=25 0.00328 0.00327 0.00256 0.00328h=30 0.00324 0.00323 0.00256 0.00323
Average 0.002785 0.00278 0.00256 0.00278Rank 3 2 1 2
MAPE h=5 112.3334 106.233 110.264 108.267h=10 107.1046 104.407 110.264 105.306h=15 104.9239 103.185 110.264 103.765h=20 106.6567 106.298 110.264 106.418h=25 104.6988 104.212 110.264 104.374h=30 103.7399 103.278 110.264 103.432
Average 106.5762167 104.6022 110.2641 105.2602Rank 3 1 4 2
Average Rank 2.67 2 3 1.67
Table 4. Model comparison for GBP/USD currency pairs
44
Tables
Error MetricsForecast Horizon
Neural Network ARIMA Mean
Forecast Ensemble
RMSE h=5 0.00799 0.00781 0.00648 0.00768h=10 0.00718 0.00687 0.00648 0.00652h=15 0.0066 0.00622 0.00648 0.00604h=20 0.00638 0.00592 0.00648 0.0059h=25 0.00606 0.00582 0.00648 0.00568h=30 0.00563 0.00543 0.00648 0.00528
Average 0.00664 0.00635 0.00648 0.00618Rank 4 2 3 1
MAE h=5 0.00766 0.00684 0.00515 0.00658h=10 0.00598 0.00558 0.00515 0.00515h=15 0.00546 0.00492 0.00515 0.00486h=20 0.00538 0.00484 0.00515 0.00488h=25 0.00511 0.00478 0.00515 0.00469h=30 0.00465 0.00441 0.00515 0.00432
Average 0.00571 0.00523 0.00515 0.00508Rank 4 3 2 1
MAPE h=5 186.672 93.456 100.229 87.6349h=10 195.671 137.262 100.229 99.9037h=15 162.704 114.257 100.229 99.798h=20 148.524 111.564 100.229 100.037h=25 141.489 109.427 100.229 100.364h=30 134.069 108.4 100.229 100.239
Average 161.5213 112.3942 100.2289 97.99607Rank 4 3 2 1
Average Rank 4 2.67 2.33 1
Table 5. Model comparison for JPY/USD currency pairs
45
Tables
Error MetricsForecast Horizon
Neural Network ARIMA Mean
Forecast Ensemble
RMSE h=5 0.00907 0.00913 0.00709 0.00907h=10 0.00709 0.00712 0.00709 0.00709h=15 0.00599 0.00602 0.00709 0.00599h=20 0.00872 0.00874 0.00709 0.00872h=25 0.00856 0.00857 0.00709 0.00856h=30 0.00863 0.00864 0.00709 0.00863
Average 0.00801 0.00804 0.00709 0.00801Rank 2 3 1 2
MAE h=5 0.00815 0.00826 0.00595 0.00815h=10 0.00595 0.006 0.00595 0.00595h=15 0.00474 0.00478 0.00595 0.00474h=20 0.006 0.00602 0.00595 0.006h=25 0.00599 0.00601 0.00595 0.00599h=30 0.00629 0.00631 0.00595 0.00629
Average 0.00619 0.00623 0.00595 0.00619Rank 2 3 1 2
MAPE h=5 101.572 104.116 113.129 101.572h=10 110.984 113.915 113.129 110.984h=15 106.981 108.878 113.129 106.981h=20 103.209 104.278 113.129 103.209h=25 102.724 103.607 113.129 102.724h=30 102.168 102.885 113.129 102.168
Average 104.6062 106.2796 113.1294 104.6062Rank 1 2 3 1
Average Rank 1.67 2.67 1.67 1.67
Table 6. Calculation of economic edge
46
Tables
Worst Performing model
Ensemble Economic Edge (Ensemble vs. worst model)
Predicted EUR/USD 5897157.93(ARIMA) 5844710.08
52,448 EUR =0.052 million EUR (approx.)
Predicted GBP/USD
17760292.5 (NN) 14627348 3132944 GBP =3.13 million GBP (approx.)
Predicted JPY/USD
3687812442(ARIMA) 3611318464 76493979 JPY = 76.5 million JPY (approx.)
Assumption: Transaction of USD 1 billion
47