Hello all,
I have been using R's time series capabilities to perform analysis for quite
some time now and I am having some questions regarding its reliability. In
several cases I have had substantial disagreement between R and other packages
(such as gretl and the commercial EViews package).
I have just encountered another problem and thought I'd post it to the list.
In
this case, Gretl and EViews give me similar estimations, but R is completely
different. The EViews results and gretl results are below followed by the R
results. The model is an ARIMA(0,1,2) with a single exogenous regressor (X).
The same data set was used. Here are the estimations:
EViews:
Dependent Variable: DSPOT
Method: Least Squares
Date: 07/23/08 Time: 14:37
Sample (adjusted): 2 518
Included observations: 517 after adjustments
Convergence achieved after 8 iterations
White Heteroskedasticity-Consistent Standard Errors & Covariance
Backcast: 0 1
Variable Coefficient Std. Error t-Statistic Prob.
X(-1) 3.419048 1.185199 2.884787 0.0041
MA(1) -0.049565 0.079305 -0.624994 0.5323
MA(2) -0.249748 0.100952 -2.473914 0.0137
R-squared 0.044155
Mean dependent var 0.613926
Adjusted R-squared 0.040436
S.D. dependent var 12.36165
S.E. of regression 12.10914
Akaike info criterion 7.831584
Sum squared resid 75368.51
Schwarz criterion 7.856235
Log likelihood -2021.465
Durbin-Watson stat 1.969820
Inverted MA Roots .53 -.48
gretl:
Model 13: ARMAX estimates using the 517 observations 2-518
Estimated using Kalman filter (exact ML)
Dependent variable: (1-L) Spot
Standard errors based on Outer Products matrix
VARIABLE COEFFICIENT STDERROR T STAT P-VALUE
theta_1 -0.0491101 0.0439294 -1.118 0.26360
theta_2 -0.248075 0.0439901 -5.639 <0.00001 ***
X_1 3.40437 1.21871 2.793 0.00522 ***
Mean of dependent variable = 0.613926
Standard deviation of dep. var. = 12.3617
Mean of innovations = 0.843443
Variance of innovations = 145.801
Log-likelihood = -2021.5668
Akaike information criterion (AIC) = 4051.13
Schwarz Bayesian criterion (BIC) = 4068.13
Hannan-Quinn criterion (HQC) = 4057.79
Finally, R:
gold.data <- cbind(ts(GoldData$Spot), lag(ts(GoldData$X),-1))
gold.2 <- arima(gold.data[,1], order = c(0,1,2),
xreg=gold.data[,2], method="ML")
Call:
arima(x = gold.data[, 1], order = c(0, 1, 2), xreg = gold.data[, 2], method
"ML")
Coefficients:
ma1 ma2 gold.data[, 2]
0.019 -0.202 -2.860
s.e. 0.050 0.045 3.371
sigma^2 estimated as 148: log likelihood = -2021, aic = 4050
EViews and Gretl give comparable (and I am inclined to presume, correct)
results. R on the other hand, has the exogenous regressor with a negative
coefficient. If I use other data I encounter the same problem - agreement
between EViews and Gretl, disagreement with R (for identical data sets). Are
there any known bugs with arima estimation in R? If I use the Zelig package, I
get the same results as the arima{stats} function call. If I remove the
exogenous regressor from the estimations then I have agreement between R, Gretl
and EViews, but with the exogenous regressor (basically an ARMAX model) the
estimation results are substantially different.