DOLLAR OR PESO?

FORECASTING DOLLAR-PESO EXCHANGE RATES USING VARIOUS

FORECASTING METHODOLOGIES
An Empirical Paper
Presented to
Prof. Cesar C. Rufino
School of Economics
De La Salle UniversityManila

In Partial Fulfilment of the Requirements for

Economic Forecasting & Model Buiding (ECOFORE)
3rd Trimester, Academic Year 2013-2014

Submitted by
Tan, Daniel Joseph C.
11114371
V24

April 12, 2014

Page1

Table of Contents
I.

Introduction...................................................................................................... I2
A.

Background and Significance of the Study....................................................I2

B.

Scope and Limitations................................................................................... I2

II.

Review of Related Research and Literature.....................................................II3

A.

B.

Classical Decomposition Methods................................................................II3

1.

Additive Decomposition............................................................................II3

2.

Multiplicative Decomposition....................................................................II3
Exponential Smoothing................................................................................II3

C. ARIMA........................................................................................................... II3
D. TRAMO-SEATS for Windows (TSW) v1.1.4b rev. 194.....................................II3
III.

Results and Analysis.................................................................................... III4

A.

Additive Decomposition.............................................................................. III4

B.

Multiplicative Decomposition......................................................................III4

C. Exponential Smoothing............................................................................... III4

D. ARIMA.......................................................................................................... III4
IV.

Conclusion.................................................................................................. IV5

V.

Appendix......................................................................................................... V6
A.

Gretl Outputs............................................................................................... V6

B.

EViews Outputs............................................................................................ V6

C. TSW Outputs................................................................................................ V6
VI.

References.................................................................................................. VI7

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Introduction
A.

Page2

I.

Background and Significance of the Study

This study aims to provide suggestions on whether investors
should buy, sell, or hold onto their dollars by providing forecasts on the
monthly average peso-per-dollar exchange rate. The studys analyses
strives to evaluate the forecasting performances of various forecasting
methodologies.

B.

Scope and Limitations

The study will make use of the monthly average peso-per-USdollar data from January of 1945 to February of 2014 (830
observations). The researcher chose to forecast using monthly data to
make the ex-ante predictions more accurate to a certain degree.
Various econometric software such as Gretl, EViews, and TSW will be
used on the different forecasting methods to be discussed throughout
the study.

II.

Results and Analysis

A.

Additive Decomposition1
According to the resulting Kruskal-Wallis (KW) Statistic, there is
an apparent need to deseasonalize the data. Using the Ordinary Least
Squares (OLS) Regression, it shoes that time has a significant
correlation with monthly average peso-per-dollar exchange rate,
having a p-value that is extremely close to zero. However, the
researcher notes that the resulting forecasts have significantly high
residuals and do not fit the actual series very well. The resulting
regressions Theils U criterion, being well above the 0.55 threshold,
supports this empirical observation and further confirms OLSs
inaccuracy.

1 See the Additive (A) Decomposition sheets in AdditiveMultiplicative.xlsx

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Page3

Original vs Forecasted
LL
UL
Original
Forecasted

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Page4

To address this, the researcher opted to use the General Least

Squares (GLS) Estimation. Using the Prais-Winsten Correction, the
resulting regressions Theils U criterion has significantly decreased.
However, it still hasnt satisfied the criterions threshold requirement of
0.55 and below, but the researcher notes that the resulting forecast fits
more accurately to the actual data.2
56

95 percent interval
pes odollar
forecas t

54

52

50

48

46

44

42

40
2006

B.

2007

2008

2009

2010

2011

2012

2013

2014

2015

Multiplicative Decomposition3
The researcher observed that that despite of small quantitative
differences in the additive and multiplicative model, the Multiplicative
Decomposition has an almost identical result. Their Theils U Criterion
are both extremely above the threshold, which, after using the GLS
estimation still remain above the threshold but are significantly closer,
denoting an increase in the estimations accuracy.

2 See Gretl results in the Appendix

3 See the Multiplicative (M) Decomposition sheets in AdditiveMultiplicative.xlsx

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Page5

C.

Exponential Smoothing
Before making use of the exponential smoothing method, we
must first check for the presence of Unit Root/s. According to the
result of the Augmented Dickey-Fullet (ADF) Test 4, the monthly average
peso-per-dollar exchange rate is an integrated stochastic process of
the first degree, meaning that the exchange rate is found to be
stationary at the 1st difference.

Original vs Exponential Smoothing

Original
Exponential Smoothing

Results show that the smoothing constant, , is at 0.999 and

indicates that the most recent observation has a very significant
weight on the forecast. As seen in the graph above, the fitted values
are accurately following the actual values through time.

4 See EViews Outputs in the Appendix

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Page6

D.

ARIMA

According to TSW, the optimal ARIMA model specification for the

monthly average peso-per-dollar exchange rate is ARIMA (2,1,1)
(0,0,0). As seen in the graph below, the forecast fits well with
the actual data.
56

95 percent interval
pesodollar
forecast

54

52

50

48

46

44

42

40
2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

However, in order to comprehensively determine the accuracy of

the forecast, key statistical values must be evaluated:
MeanError2.5806e006
MeanSquaredError0.24176
RootMeanSquaredError0.49169
MeanAbsoluteError0.21502
MeanPercentageError0.21945
MeanAbsolutePercentageError1.1866
Theil'sU1.0905

Although the Mean Squared Error (MSE) and the Mean Absolute
Percentage Error (MAPE) are acceptable, the model specification
was unfortunately unable to reach the threshold requirements
for the Theils U Criterion. However, it is clearly safe to say that
despite its shortcomings, its criterion is indeed lower than the
other models.

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Conclusion

Page7

III.

Since both of the Classical Decomposition methods are similar in

that the observed time series are seasonal, their results are similar.
When compared with the Theils U criterion of the ARIMA specification
that was chosen by TSW, we could safely state that the ARIMA (2,1,1)
(0,0,0) is, in fact, the better model. However, both the Classical
Decomposition and the ARIMA models are far from being optimal for
forecasting the monthly average peso-per-dollar exchange rate.
On the other hand, the Exponential Smoothing methodology can
still be considered optimal in the sense that its forecast is relatively
accurate. However, the limitation of such accuracy is that it is only
effective in forecasting one period ahead.
In summary, the researcher has chosen ARIMA (2,1,1)(0,0,0) as
the best model for forecasting the monthly average peso-per-dollar
exchange rate. The unsuitability of all of the models used in the study
may be caused by the size of the time series as well as the sheer
number of various outliers and transitory changes present in the
dataset.

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Page8

IV.

Appendix
A.

Gretl Outputs
1.
Additive/Multiplicative GLS Regression with PraisWinsten Correction
ITERRHOESS
10.99604226.744
20.99780226.468
30.99811226.468
Model2:PraisWinsten,usingobservations1945:012014:02(T=830)
Dependentvariable:pesodollar
rho=0.998112
coefficientstd.errortratiopvalue

const1.027028.328240.12330.9019
time0.05344250.01289494.1443.76e05***
Statisticsbasedontherhodifferenceddata:
Meandependentvar18.28770S.D.dependentvar17.98124
Sumsquaredresid226.4545S.E.ofregression0.522968
Rsquared0.999156AdjustedRsquared0.999155
F(1,828)5.631535Pvalue(F)0.017868
rho0.328240DurbinWatson1.343363

2.

Additive/Multiplicative GLS Forecast

56

95 percent interval
pesodollar
forecast

54

52

50

48

46

44

42

40
2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Page9

3.

ARIMA Specification

Functionevaluations:54
Evaluationsofgradient:12
Model3:ARMAX,usingobservations1945:022014:02(T=829)
EstimatedusingKalmanfilter(exactML)
Dependentvariable:(1L)pesodollar
StandarderrorsbasedonOuterProductsmatrix
coefficientstd.errorzpvalue

const0.01446570.04822120.30000.7642
phi_10.1201990.2973260.40430.6860
phi_20.07318440.1145200.63910.5228
theta_10.4766260.2930631.6260.1039
time8.92853e050.0001004700.88870.3742
Meandependentvar0.051743S.D.dependentvar0.523173
Meanofinnovations2.58e06S.D.ofinnovations0.491691
Loglikelihood587.8543Akaikecriterion1187.709
Schwarzcriterion1216.030HannanQuinn1198.570
RealImaginaryModulusFrequency

AR
Root12.96540.00002.96540.5000
Root24.60780.00004.60780.0000
MA
Root12.09810.00002.09810.5000

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Page10

For95%confidenceintervals,t(828,0.025)=1.963
pesodollarpredictionstd.error95%interval
2005:1154.56057955.727684
2005:1253.61175054.582829
2006:0152.61713653.635893
2006:0251.81680052.643258
2006:0351.21891351.844533
2006:0451.35972251.247876
2006:0552.12727351.388520
2006:0653.15666752.154723
2006:0752.39760053.182274
2006:0851.36177352.424741
2006:0950.40121151.390971
2006:1050.00381050.432323
2006:1149.84342950.035773
2006:1249.46700049.875796
2007:0148.91431849.500179
2007:0248.38135048.948641
2007:0348.51722748.416780
2007:0447.82244448.552502
2007:0546.81414347.859132
2007:0646.16015046.852835
2007:0745.62509146.200178
2007:0846.07371445.666230
2007:0946.13145046.114107
2007:1044.38028646.171835
2007:1143.21763244.424077
2007:1241.74288943.263719
2008:0140.93809141.791862
2008:0240.67110040.988684
2008:0341.25236840.722298
2008:0441.81957141.302570
2008:0542.90195241.868803
2008:0644.28055042.949241
2008:0744.95626144.325337
2008:0844.87694744.999873
2008:0946.69222744.920810
2008:1048.02522746.732764
2008:1149.18620048.063348
2008:1248.09423549.222230
2009:0147.20715048.132428
2009:0247.58455047.247118
2009:0348.45804547.623906
2009:0448.21652648.495854
2009:0547.52415048.254891
2009:0647.90528647.563923
2009:0748.14600047.944440
2009:0848.16072248.184801
2009:0948.13940048.199596
2009:1046.85127348.178415
2009:1147.03242146.892821
2009:1246.42105347.073728
2010:0146.02760046.463615
2010:0246.31160046.071006
2010:0345.74247846.354570
2010:0444.62657945.786624
2010:0545.59710544.672933
2010:0646.30270045.641727
2010:0746.32027346.346091
2010:0845.18247646.363731
2010:0944.31366745.228184
2010:1043.44490044.361116
2010:1143.49184243.494090
2010:1243.95475043.541044
2011:0144.17219044.003179
2011:0243.70310044.220310
2011:0343.51600043.752206
2011:0443.24015843.565560
2011:0543.13072743.290340
2011:0643.36566743.181217
2011:0742.80876243.415813
2011:0842.42090542.860061
2011:0943.02563642.473037
2011:1043.45140043.076728
2011:1143.27452643.501789
2011:1243.64857143.325350
2012:0143.61914343.698790
2012:0242.66076243.669517
2012:0342.85736442.713047
2012:0442.69977842.909378
2012:0542.85145542.752191
2012:0642.77645042.903682
2012:0741.90540942.828920
2012:0842.04521141.959625
2012:0941.74895042.099263
2012:1041.45213641.803663
2012:1141.12215841.507510
2012:1241.00672241.178256
2013:0140.72950041.063139
2013:0240.67225040.786541
2013:0340.71273740.729500
2013:0441.14223840.770011
2013:0541.29757141.198803
2013:0642.90694741.353944
2013:0743.35587042.960382
2013:0843.86394143.408557
2013:0943.83181043.915771
2013:1043.18247643.883801
2013:1143.55460043.235794
2013:1244.10430043.607316
2014:0144.92657144.156079
2014:0244.89500044.976899
2014:0344.9454880.52296843.91898945.971988
2014:0444.9959820.73889143.54566246.446302
2014:0545.0464810.90410043.27188546.821078
2014:0645.0969861.04298043.04979047.144183
2014:0745.1474971.16498942.86081747.434177
2014:0845.1980131.27498142.69543847.700587
2014:0945.2485341.37584142.54798747.949081
2014:1045.2990611.46945442.41476948.183354
2014:1145.3495941.55712742.29321448.405974
2014:1245.4001321.63981542.18144948.618815
2015:0145.4506761.71823942.07805948.823292
2015:0245.5012251.79295941.98194549.020505

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

EViews Outputs
1.

Page11

B.

Exponential Smoothing ADF Test

Null Hypothesis: D(PESODOLLAR) has a unit root
Exogenous: Constant
Lag Length: 4 (Fixed)

Augmented Dickey-Fuller test statistic

Test critical values:
1% level
5% level
10% level

t-Statistic

Prob.*

-10.92992
-3.438061
-2.864833
-2.568578

0.0000

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

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(PESODOLLAR,2)
Method: Least Squares
Date: 04/12/14 Time: 23:00
Sample (adjusted): 1945M07 2014M02
Included observations: 824 after adjustments
Variable

Coefficient

Std. Error

t-Statistic

Prob.

D(PESODOLLAR(-1))
D(PESODOLLAR(-1),2)
D(PESODOLLAR(-2),2)
D(PESODOLLAR(-3),2)
D(PESODOLLAR(-4),2)
C

-0.626088
-0.019879
-0.117240
-0.022569
-0.040193
0.032655

0.057282
0.053957
0.048990
0.041691
0.035062
0.017453

-10.92992
-0.368433
-2.393144
-0.541339
-1.146353
1.870996

0.0000
0.7126
0.0169
0.5884
0.2520
0.0617

R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
F-statistic
Prob(F-statistic)

0.345088
0.341084
0.493940
199.5728
-584.9928
86.20438
0.000000

Mean dependent var

S.D. dependent var
Akaike info criterion
Schwarz criterion
Hannan-Quinn criter.
Durbin-Watson stat

-3.83E-05
0.608498
1.434449
1.468775
1.447617
2.004324

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions

Simple Exponential Smoothing

Page12

2.

Original vs Exponential Smoothing

Original
Exponential Smoothing

Date: 04/12/14 Time: 23:03

Sample: 1945M01 2014M02
Included observations: 830
Method: Single Exponential
Original Series: PESODOLLAR
Forecast Series: PESODOSM
Parameters:
Alpha
Sum of Squared Residuals
Root Mean Squared Error
End of Period Levels:

V.

0.9990
232.4297
0.529184
Mean

44.89503

References
Gujarati, D. N., & Porter, D. C. (2004). Basic Econometrics. Singapore:
The McGraw-Hill Companies.

Rufino, C. C. (2012). Signal Extraction from the Philippine National Accounts

Statistics using ARIMA Model-based Methodology. DLSU Business &
Economics Review 21.2, 1-16.

An Analysis on the Philippine GDP and its Determinants in Carbon Dioxide Emissions