International Journal of Probability and Statistics 2019, 8(2): 30-36

DOI: 10.5923/j.ijps.20190802.02

Exchange Rate Forecasting Using Non-linear Threshold

Models
Akintunde Mutairu Oyewale1,*, Chigozie Kelechi Acha2, Agunloye Oluokun Kasali3,
Eriobu Nkiru Obioma4, Oyekunle Janet Olufunmike1, Abdulazeez Ismail Adeyinka1

1
Department of Statistics, School of Applied Sciences the Federal Polytechnic, Ede, Osun State, Nigeria
2
Michael Okpara University of Agriculture, Umudike, Abia State, Nigeria
3
Department of Mathematics, Faculty of Science Obafemi Awolowo University, Ile-Ife, Osun State, Nigeria
4
Department of Statistics, Faculty of Physical Science, Nnamdi Azikiwe University, Awka, Anambra State

Abstract Foreign exchange is one of the most important financial instruments very volatile and chaotic in nature.
Nowadays, the role of the foreign exchange market is becoming more and more important in the financial markets around the
world; it remains the only instruments worldwide to measure the standard of living, Economic performance and country
standing among the committee of nations. The foreign exchange market which is an over-the-counter market is used for the
trading of currencies. It makes the foreign exchange market the largest and most liquid market among the financial markets.
Necessary mathematical frame work was put in place, this was illustrated with data from record of Central Bank of Nigeria
through their official website. Through this the performance of nonlinear threshold models in forecasting the exchange rate of
Nigeria in relation to United States of American dollar as bench mark. The software used for the analysis was Econometrics
View (E-view). Stationarity tests were carried out before the analysis, (the original data was not stationary, but at first
difference it was stationary, thereafter comprehensive data analysis was performed). The forecasting results indicate that
SETAR models did not outperform Random Walk in any period. TAR models offered promising results in the period. This
study supports the general belief that the exchange rates are chaotic, volatile and very difficult to forecast and this applies to
Exchange rate system of Nigeria as it is the case in the developed world.
Keywords Forecasting, Stationarity test, Exchange rate, Nonlinearity, SETAR, TAR

include ARIMA model (Tseng, 2001), [19], Least Squared

1. Introduction model [10] or Purchasing Power Parity model and [3]. [13]
used ARIMA model for forecasting inflation in Irish, [14]
Foreign exchange is one of the most important financial used Arima model for forecasting stock price. ARIMA is
instruments. Nowadays, the role of the foreign exchange also used for predicting stock price in the research of [12],
market is becoming more and more important in the financial [11]. It is also used for forecasting the price of gold [8].
markets around the world. The foreign exchange market According to [1] Triennial Survey, turnover in global
which is an over-the-counter market is used for the trading of foreign exchange markets averaged $5.1 trillion per day.
currencies. The trading is happening 24 hours a day around This is down from $5.4 trillion in April 2013, a month which
the world and a great number of currencies is being had seen heightened activity in Japanese yen against the
transacted every hour. It makes the foreign exchange market background of monetary policy developments at that time. In
the largest and most liquid market among the financial addition, exchange rate movements are used to compare
markets. influence of the previous surveys on the current year. In
There are so many studies on exchange rate forecasting particular, the appreciation of the US dollar between 2013
carried out worldwide. Until now, there are a lot of compared with 2016. When valued at constant (April 2016)
forecasting models, each having its point of strengths and exchange rates, turnover increased slightly, by about 4%
weaknesses used in exchange rate forecasting, such study between April 2016 and April 2013. Nevertheless, the latest
developments contrast with the strong growth in turnover
* Corresponding author: observed between Triennial Surveys since 2001.
[email protected] (Akintunde Mutairu Oyewale)
Published online at http://journal.sapub.org/ijps
The foreign exchange rate has two main uses. Firstly, it
Copyright © 2019 The Author(s). Published by Scientific & Academic Publishing allows the businesses to exchange its currency to a target
This work is licensed under the Creative Commons Attribution International currency in the determined foreign exchange rate. Thus, it
License (CC BY). http://creativecommons.org/licenses/by/4.0/ benefits the global trade and investment. Secondly, it also
 International Journal of Probability and Statistics 2019, 8(2): 30-36 31

provides the speculation and expedites the carry trade, in also exists high risk in the speculation.
which there are substantial profits available. However, there

Figure 1. Foreign Echange Market Turnover by Currency Pairs

The figure 1 above summarizes the exchange rate relative is gradually becoming a more realistic representation of data
to some selected currencies in the world market. generation processes. In finance, for instance, stock returns
tend to be more correlated when there is low volatility than
when volatility is high. A similar behavior has been observed
2. Review of Relevant Literature in exchange rate mechanisms where the exchange rate may
be constrained to lie within a pre-defined target zone [4]. To
So many Empirical econometric modeling works in accommodate this kind of dynamic behavior using time
Agricultural Economics, finance and economic data assume series data, regime-switching models (RSM) have been
that relationships are linear. Economic theory plays a introduced ([15] & [16]; [7]). Threshold autoregressive
passive role on this issue, and thus most applied research (TAR) model begins to be regularly appears in the
finds it convenient to assume linearity. In the 2000's, some agricultural economics literature as a model that is popularly
researchers try to challenge the empirical theory ―the used [18], and extensively discussed in Tong [17].
random walk model predicts the exchange rate best‖. They
used different methods and chose different models. For
examples, data from Central and Eastern European countries 3. Mathematical Preliminaries of TAR
was used to compare the forecasting models in transition
economies [2]. Intraday foreign exchange rates were used as
and SETAR Models
observations [9]. They found that some sophisticated time 3.1. SETAR Models
series models such as the Markov regime-switching model
have better performance than the random walk models under The SETAR model is a convenient way to specify a TAR
the condition of intensive time period. Thus, the empirical model because qt is defined simply as the dependent variable
theory ―the random walk model predicts the exchange yt . In this case, the process can be formally written as
rate best‖ does not work well sometimes. The thought 0,1  1,1 yt 1  2,1 yt 2   p1,1 yt  p1   t if yt 1  c
expressed began to change based on irregularities observed yt  
in economic and financial data, that non-linear specifications  0,2  1,2 yt 1  2,2 yt 2   p 2,1 yt  p 2   t if yt 1  c
32 Akintunde Mutairu Oyewale et al.: Exchange Rate Forecasting Using Non-linear Threshold Models

It is interesting to highlight that the estimation of SETAR The estimation of the threshold values in Step 3 requires
models requires the application of least squares procedures that each regime contains enough observations for reliable
only, more specifically, sequential conditional least squares. estimation of the AR coefficients. About fifteen percent
For the two-regime SETAR model, the steps can be outlined (15%) of the observations on each regime seems to work
as follows: well.
Step 1: Set P1  P2  P for simplicity and estimate the
3.2. TAR Models
AR coefficients conditional on the value of the threshold (c).
Step 2: Calculate conditional residuals ˆt and estimated  i ut 1   t if t 1  0
ut  
variances ˆ t from the coefficients in step 1.  2ut 1   t if t 1  0

Step 3: Obtain least squares estimates of c by minimizing If the above sequence is stationary, the least squares
the residual variance ˆ  c  over all possible values of the estimates of P1 and P2 have an asymptotic multivariate
threshold coefficient c, that is, normal distribution. The process is formally specified as:
Cˆ  arg min cˆ 2  c  ut  It i ut 1   I  It  2ut 1   t

This minimization requires a direct search over the 1 if u1  0

It  
ordered values of yt [6] provides an approach free of 0 if u1  0
nuisance parameters that are involved in estimation. Also
note that in more complex models, the direct search may be
over values of c and d, and thus, the estimate of the variance 4. Description and Data Anaysis
in Step 3 would be conditional on values of c and d (e.g., [6]).
Step 4: Obtain final estimates of ˆ and ˆ 2 from results The exchange rate data used in the study was extracted
from Central Bank of Nigeria Bulletin of 2017. The data
in Step 3.
was obtained for 1996 to 2018.
ORIGINAL DATA
400

350

300

250

200

150

100

50

0
96 98 00 02 04 06 08 10 12 14 16 18

Figure 2. Line Graph of the Leveled Exchange Rate of Naira/ Dollar

first diff
80

60

40

20

-20

-40
96 98 00 02 04 06 08 10 12 14 16 18

Figure 3. Line Graph of the First Difference Exchange Rate of Naira/Dollar

 International Journal of Probability and Statistics 2019, 8(2): 30-36 33

4.1. Identification of a Stationary Condition of the Series The results on the correlogram of the leveled for the
The line graph of the series Figure 2 indicates the series shows stronger evidence of non-stationarity since its
non-stationarity of the series. There is evidence of volatility autocorrelation coefficient function (ACF) of the residuals
as the values do not fluctuate around a constant mean. The does not quickly decay to zero. On the other hand, the
first differences of the series were taken (figure 3) and the correlogram of the first difference shows that it is consistent
graphs seem to fluctuate around a constant mean of zero with mean stationarity because most of the values promptly
value. decay to zero. Tables 1 and 2 below show both the
correlgrams for level and first difference.

Table 1. Correlogram for Level

Date: 11/10/19 Time: 17:20

Sample: 1996M01 2018M12
Included observations: 276
Autocorrelation Partial Correlation AC PAC Q-Stat Prob
.|******* .|******* 1 0.983 0.983 269.50 0.000
.|******* *|. | 2 0.962 -0.110 528.75 0.000
.|******* .|. | 3 0.940 -0.061 776.84 0.000
.|******* .|. | 4 0.917 0.001 1014.0 0.000
.|******| .|. | 5 0.895 0.019 1241.0 0.000
.|******| *|. | 6 0.872 -0.069 1457.1 0.000
.|******| .|. | 7 0.851 0.059 1663.6 0.000
.|******| *|. | 8 0.828 -0.067 1859.9 0.000
.|******| .|. | 9 0.805 -0.011 2046.3 0.000
.|******| .|. | 10 0.782 -0.014 2222.9 0.000
.|***** | .|. | 11 0.759 -0.011 2389.8 0.000
.|***** | .|. | 12 0.736 -0.023 2547.4 0.000
.|***** | .|. | 13 0.713 0.000 2695.7 0.000
.|***** | .|. | 14 0.690 -0.036 2835.0 0.000
.|***** | .|. | 15 0.666 0.001 2965.6 0.000
.|***** | .|. | 16 0.644 -0.006 3087.8 0.000
.|**** | .|. | 17 0.619 -0.054 3201.4 0.000
.|**** | .|. | 18 0.595 -0.019 3306.6 0.000
.|**** | .|. | 19 0.569 -0.053 3403.3 0.000
.|**** | .|. | 20 0.546 0.072 3492.5 0.000
.|**** | .|. | 21 0.522 -0.037 3574.5 0.000
.|**** | .|. | 22 0.498 -0.019 3649.5 0.000
.|*** | .|. | 23 0.475 -0.015 3717.9 0.000
.|*** | .|. | 24 0.451 -0.015 3779.9 0.000
.|*** | .|. | 25 0.428 -0.002 3836.0 0.000
.|*** | .|. | 26 0.405 -0.024 3886.3 0.000
.|*** | .|. | 27 0.381 -0.022 3931.1 0.000
.|*** | .|. | 28 0.358 -0.022 3970.7 0.000
.|** | .|. | 29 0.333 -0.045 4005.1 0.000
.|** | .|. | 30 0.310 0.048 4035.2 0.000
.|** | .|. | 31 0.289 0.008 4061.4 0.000
.|** | .|* | 32 0.274 0.176 4085.0 0.000
.|** | .|. | 33 0.263 0.045 4106.8 0.000
.|** | .|. | 34 0.251 -0.037 4126.8 0.000
.|** | .|. | 35 0.240 -0.023 4145.1 0.000
.|** | .|. | 36 0.228 0.001 4161.7 0.000
34 Akintunde Mutairu Oyewale et al.: Exchange Rate Forecasting Using Non-linear Threshold Models

Table 2. Correlogram for First Difference

Date: 11/10/19 Time: 17:15
Sample: 1996M01 2018M12
Included observations: 275
Autocorrelation Partial Correlation AC PAC Q-Stat Prob
.|. | .|. | 1 0.069 0.069 1.3351 0.248
.|* | .|* | 2 0.091 0.086 3.6266 0.163
.|. | .|. | 3 -0.007 -0.019 3.6400 0.303
.|. | .|. | 4 0.032 0.026 3.9241 0.416
.|. | .|. | 5 -0.062 -0.064 5.0043 0.415
.|. | .|. | 6 0.012 0.016 5.0478 0.538
.|. | .|. | 7 0.017 0.028 5.1342 0.644
.|. | .|. | 8 -0.019 -0.028 5.2423 0.731
.|. | .|. | 9 -0.008 -0.004 5.2599 0.811
.|. | .|. | 10 -0.004 -0.003 5.2637 0.873
.|. | .|. | 11 -0.029 -0.028 5.5060 0.904
.|. | .|* | 12 0.066 0.076 6.7682 0.873
.|. | .|. | 13 0.064 0.057 7.9414 0.847
*|. | *|. | 14 -0.105 -0.130 11.145 0.675
.|* | .|* | 15 0.075 0.089 12.785 0.619
.|* | .|* | 16 0.158 0.169 20.126 0.215
.|. | .|. | 17 0.021 -0.018 20.251 0.262
.|. | .|. | 18 0.020 0.005 20.372 0.312
.|. | .|. | 19 0.008 -0.015 20.390 0.371
.|. | .|. | 20 0.012 0.013 20.433 0.431
.|. | .|. | 21 -0.016 0.016 20.506 0.489
.|. | *|. | 22 -0.050 -0.071 21.252 0.505
.|. | .|. | 23 0.027 0.035 21.467 0.553
*|. | *|. | 24 -0.086 -0.078 23.688 0.480
.|* | .|. | 25 0.075 0.071 25.398 0.440
.|. | .|. | 26 -0.042 -0.012 25.945 0.466
.|. | .|. | 27 0.009 -0.003 25.970 0.520
.|. | .|. | 28 0.045 0.018 26.589 0.541
*|. | *|. | 29 -0.093 -0.124 29.262 0.451
*|. | *|. | 30 -0.152 -0.110 36.444 0.194
*|. | *|. | 31 -0.080 -0.068 38.457 0.168
.|. | .|. | 32 0.006 0.006 38.468 0.200
.|. | .|. | 33 -0.021 -0.011 38.608 0.231
.|. | .|. | 34 0.002 0.001 38.609 0.269
.|. | .|. | 35 0.008 -0.003 38.631 0.309
.|. | .|. | 36 -0.003 -0.011 38.634 0.351

The stationary conditions of the series were formally Table 3. Original Data
verified by using Unit Root test (URT) for the leveled Null Hypothesis: VALUES has a unit root
and first differences of the series. We tested for a unit root Exogenous: Constant
using the augmented Dickey-Fuller (ADF) statistic. At level, Lag Length: 0 (Automatic - based on SIC, maxlag=15)
(table 3) all the series are not stationary but at first difference t-Statistic Prob.*
(table 4) all series appeared stationary as shown in the tables
Augmented Dickey-Fuller test statistic 0.117358 0.9666
below:-
Test critical values: 1% level -3.453997
5% level -2.871845
10% level -2.572334

*MacKinnon (1996) one-sided p-values.

 International Journal of Probability and Statistics 2019, 8(2): 30-36 35

Table 4. First Difference Panel A (table 6) is regime 1 of the two-regime SETAR and
Null Hypothesis: D(VALUES) has a unit root panel B (table 7) is regime 2. We select the most appropriate
Exogenous: Constant model by minimizing value of AIC. We build up a
Lag Length: 0 (Automatic - based on SIC, maxlag=15) two-regime SETAR model, SETAR (2, 7, 4).
t-Statistic Prob.* Table 7. Panel B Regime 2 Observations Included 308
Augmented Dickey-Fuller test statistic -14.89460 0.0000
RANK MODELS AIC
Test critical values: 1% level -3.454085
1 AR(4) 9.0982
5% level -2.871883
2 AR(10) 9.0021
10% level -2.572354
3 AR(3) 9.0882
*MacKinnon (1996) one-sided p-values. 4 AR(9) 9.0071
5 AR(7) 9.0765
4.2. Results of ARIMA Model Selection
6 AR(4) 9.0974
To identify the appropriate model parameter of ARIMA
7 AR(5) 9.0053
(p, d, q), AIC is adopted. From the unit root test, we obtain d
to be 1. As to parameters p and q, we run the regression 8 AR(1) 9.0642
through the combinations of from p = 1 to p = 10 and from 9 AR(2) 9.0523
q = 1 to q = 10. For sake of saving space, we just list the top 10 AR(6) 9.7432
10 with higher AIC values here. As shown in Table 5 below
the model with the smallest value (9.14334) of AIC is the 4.4. Results Out-of-sample Forecasting Performance
optimal ARIMA (6, 1, 9) chosen.
In this section, we check out-of-sample forecast
Table 5. Arima Model Results performance of the two models. By checking multi-criteria
RANK MODELS AIC
(MSE, MAE, AMPE, and MAPE) on ARIMA and SETAR,
we can compare the residual of these two models. For the
1 ARIMA ( 6,1, 9) 9.14340
mean squared error (MSE), SETAR is smaller than ARIMA
2 ARIMA ( 9,1,7) 9.14403 with 4479.28 and 5744.65. This means SETAR has a fewer
3 ARIMA ( 9,1,9) 9.14545 errors in the standard of MSE. For the mean absolute error
4 ARIMA ( 9,1,6) 9.14505 (MAE), the adjusted mean absolute property error (AMAPE)
5 ARIMA ( 8,1,9) 9.14871 and mean absolute property error (MAPE), SETAR is also
6 ARIMA ( 10,1,9) 9.14879 smaller than ARIMA with fewer errors. According to our
7 ARIMA ( 6,1,6) 9.14978
result shown in the table 8 below, the SETAR model is better
than the ARIMA model over the sample period (second
8 ARIMA ( 9,1,8) 9.15125
regime).
9 ARIMA ( 6,1,5) 9.15178
Comparison of Forecasting Power.
10 ARIMA ( 6,1,10) 9.15368
Table 8. Comarison of Forecasting Power by Models
4.3. Results of SETAR Model MSE MAE AMAPE(%) MAPE (%)
Table 6. Panel A Regime 1 Observations Included 764 SETAR 4479.28 45.51 90.68 89.02
ARIMA 5744,65 58.32 304.20 190.91
RANK MODELS AIC
1 AR(7) 9.8432
2 AR(9) 9.8120 5. Conclusions
3 AR(1) 9.8743
4 AR(6) 9.8321
At first, we believed there is no way for a linear regression
to suit a series forever where stock prices follow a non-linear
5 AR(3) 9.8845
trend. Due to the economic environment changing, the stock
6 AR(4) 9.8764
market will be affected and change over time. Therefore,
7 AR(2) 9.8221 non-linear regression should be better than linear regression
8 AR(10) 9.8765 in the exchange rate market. First step of handling time series
9 AR(8) 9.9872 data is to check stationary state in the mean. We found out
10 AR(5) 9.9721 there was a unit root existed so we analyzed the
first-difference. Next, we constructed an ARIMA model
In our self-exciting threshold autoregressive (SETAR) by using AIC selection criteria. And we build up a SETAR
model, we assume that a variable Naira is a linear model by AIC as well. Afterward, we checked the
autoregression within a regime. As there are two regimes in forecasting power by four criteria (MSE, MAE, AMAPE,
the study, the model could be written as SETAR (2, p, p). and MAPE) and all of those standards showed that SETAR
36 Akintunde Mutairu Oyewale et al.: Exchange Rate Forecasting Using Non-linear Threshold Models

has a stronger predicting power than ARIMA. The results [10] Hongxing, L., Zhaoben, F., & Dongming, Z. (2007).
also support previous assumptions of this thesis. Non-linear GBP/USD Currency Exchange Rate Time Series Forecasting
Using Regularized Least-Squares Regression Model.
SETAR model is better than linear ARIMA model in Nigeria Proceeding of the World Congresson Engineering, Vol 2.
exchange rate market.
[11] Isenah, G. M. O., Olusanya E. (2014). Forecasting Nigerian
Stock Market Returns Using ARIMA and Artificial Neural
Network Models. CBN Journal of Applied Statistics, Vol.5
(No.2).
REFERENCES [12] Jarrett, J. E. K., Eric (2011). Arima modeling with
intervention to forecast and analyze Chinese stock prices.
[1] Bank for International Settlements 2016. Triennial Central
International Journal of Engineering Business Management,
Bank Survey of Foreign Exchange and Derivatives Market
Vol.3.
Activity in April 2010: Preliminary Global Results. Bank for
International Settlement, September. [13] Meyler, A., Kenny, G., & Quinn, T. (1998). Forecasting Irish
Inflation Using Arima Model (Vol. Vol. 1998, pp. 1-48):
[2] Cuaresma, J. C., B. _Egert, & R. MacDonald (2005):
Ireland: Central Bank and Financial Services.
\Non-linear exchangerate dynamics in target zones: A bumpy
road towards a honeymoon."Working Paper 771, William [14] Mondal, P. S., Labani & Goswami, Saptarsi (2014). Study on
Davidson Institute. Effectiveness of Time Series Modelling (Arima) in
Forecasting Stock Prices. International Journal of Computer
[3] David, H., Lee, J., & Takizawa, H. (2010). In which exchange
Science, Vol4 (No.2).
rate models do forecastters trust?: IMF Working Paper.
[15] Priestly, E. (1980) “Using Threshold Cointegration to
[4] Franses, P. H. and D van Dijk. Non-linear time series models
Estimate Asymmetric Price Transmission in the Swiss Pork
in empirical economics. NewYork, Cambridge University
Market,” Applied Economics: 679-87.
Press, 2000.
[16] Priestly, E. (1988): “Threshold-Autoregressive,
[5] Hansen, B. E. (1996): “Inference when a Nuisance Parameter
Median-Unbiased, and Cointegration Tests of Purchasing
is not Identified under the Null Hypothesis,” Econometrica,
Power Parity,” International Journal of Forecasting, 14,
64, 413-430.
171-186.
[6] Hansen, B. E. (2000): “Sample Splitting and Threshold
[17] Tong, H. (1990) Threshold Models in Non-Linear Time
Estimation,” Econometrica, 68, 575-603.
Series Analysis. New York: Springer Verlag.
[7] Granger, C. W. J., and Terasvirta, T. (1993): Modeling
[18] Tong, H., & Lim, K. S. (1980). Threshold autoregressive,
Nonlinear Economic Relationships, Oxford, U.K. Oxford
limit cycles and cyclical data. Journal of the Royal Statistical
University Press.
Society Series B, 42(3), 245–292.
[8] Guha, B., & Bandyopadhyay, G. (2016). Gold Price
[19] Znaczko, T. M. (2013). Forecasting Foreign Exchange Rates.
Forecasting Using Arima Model. Journal of Advanced
Applied Economics Theses, Vol.4.
Management Science, Vol.4 (No.2).
[9] Hong, Y. and Lee, T.H., (2003). Inference on predictability
of foreign exchange rates via generalized spectrum and
non-linear time series models. Review of Economics and
Statistics, 85, 1048-1062.