I want to study a time series using ARIMA. To understand what is the best ARIMA model, I use the function
Arima (dataset, order, seasonal) from the package
R. You can see dataset I am trying to analyse below.
TEST <- c(0, 0, 0, 8.718396e-04, 6.345178e-04, 4.752852e-04, 7.390983e-04, 2.973536e-04, 0, 0, 0, 0, 0, 1.138822e-04, 9.984026e-05, 0, 0, 7.015575e-05, 6.248828e-05, 0, 0, 8.847600e-05, 7.894841e-05, 3.533444e-05, 3.162255e-05, 2.840909e-05, 2.558068e-05, 2.313048e-05, 2.099782e-05, 0, 0, 0, 0, 0, 1.159205e-05, 1.077714e-05, 0, 0)
Arima(TEST, c = order(0,1,1), seasonal = FALSE) in
R, I obtained the following results:
Series: TEST ARIMA(0,1,1) Coefficients: ma1 -0.2155 s.e. 0.1741 sigma^2 estimated as 3.221e-08: log likelihood=274.33 AIC=-544.67 AICc=-544.33 BIC=-541.39
However, if I try to load the same dataset in
Analyze > Forecasting > Create Traditional Models... and I select
ARIMA(0,1,1), my results are different and, in particular, the sign of my
ma1 is opposite, i.e.
ma1 = .215 SE = .163
I am a newbie in time trend analysis, and I performed these calculations for multiple datasets, some of them very poor. Initially, I associated my sign differences to the low quality of the dataset, however I am not sure if this could be the case also in this situation, given that the dataset does not seem so bad to justify radically difference results using different programs and, consequently, different algorithms.
I also tried to compare the plots obtained from
R and from
SPSS, and at first glance the seem identical.
Finally, I did some tests also looking at the autoregressive part, and I did not notice a similar phenomenon of sign inversion.
For example, if I use the same dataset and an
ARIMA (1,1,0) I get the following results in
Coefficients: ar1 -0.1549 s.e. 0.1580 sigma^2 estimated as 3.26e-08: log likelihood=274.11 AIC=-544.23 AICc=-543.89 BIC=-540.95
ar1 = -.155 s.e. = .165
What could be the reason for this difference? And which results should I trust?