I've sampled 100 variables from a Gauss distribution with mean 0 and standard deviation 1.

> set.seed(1)
> wn=rnorm(100)

Then I've fitted an AR(1) model with the arima command and sent the results to the wnF variable

> wnF=arima(wn, order=c(1,0,0))

Finally I've requested the estimated coefficients

> wnF$coef
         ar1    intercept 
-0.003655755  0.108935363 

Now I want to replicate the computation R. I exported the data to an Excel file (https://www.dropbox.com/s/6c8ukcbtxe9gqp1/DataTester.xlsx?dl=0) and computed the following model:

$$ x{_t}-\mu = \psi_1x_{t-1}+\omega_t $$

I've replaced $\mu$ and $\psi_1$ with the intercept and ar1 coefficients reported by the wnF$coef command. I've also replaced $\omega_t$ with zero, since I've sampled the data from a zero-mean population.

Finally I've compared the residuals from the R computed model (wnF$residuals) with the residuals I've computed in the Excel file and I've noticed that they differ about $\delta<0.0005$ in absolute value.

I know that 0.0005 is not much, but when dealing with such small values it may not be negligible.

I also find strange that there is no difference up the fourth decimal place.

Can you please help me finding the origin of the difference?

  • 1
    $\begingroup$ You should have $x_t-\mu=\psi_1(x_{t-1}-\mu)+w_t$, that is, you should include the mean not only on the left hand side but also on the right hand side. Also, $w_t$ need not be zero, or actually, it should not be zero. $\endgroup$ – Richard Hardy Dec 1 '15 at 11:11
  • $\begingroup$ Subtracting the mean on the right side greatly improved the model. The difference between R computed residuals and Excel computed residuals is now less than 0.000001. Regarding $\omega_t$, I really don't know how to estimate it. $\endgroup$ – Eduardo Dec 1 '15 at 11:28
  • $\begingroup$ @Eduardo if you wish to estimate an AR(q) model in R, why not just use the LM command? It is much more (IMO) intuitive than the arima-command. $\endgroup$ – Repmat Dec 1 '15 at 12:28
  • 2
    $\begingroup$ @Repmat the estimates of AR coefficients produced by lm (what I assume you mean by "LM") are not maximum likelihood estimates and will not be the same as those produced by arima. $\endgroup$ – Glen_b -Reinstate Monica Dec 1 '15 at 12:52

After some research I found out that the difference between R computed residuals given by wnF$residuals and the residuals I've computed externaly in Excel file, was originated from the lack of precision of the data passed to Excel.

At first, I had passed data to Excel with only 7 decimal places. After repeating the procedure with 15 decimal places, the difference almost disapeared.

Also, as Richard Hardy commented, the model I was fitting was not correct. The correct model is:

$$ x_t-\mu=\psi_1(x_{t-1}-\mu)+\omega_t $$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.