# [R-SIG-Finance] ARIMA,GARCH and differences

babel at centrum.sk babel at centrum.sk
Sun Apr 12 20:01:32 CEST 2009

```Thank you for your reply, you are right with that logaritmisation, I fully agree, but the problem is this.  The I in ARIMA makes differences for you. It is obvious from the output, because the AR coefficients are very similar in the case, when I use the non stationary price series and ARMA(5,1,0) and when I use return series (lets say stationary) and model ARMA(5,0) , but NOT THE SAME. Please take o look to the following output:

Call:
armaFit(formula = ~arma(5, 0), data = ret)
Model:
ARIMA(5,0,0) with method: CSS-ML

Coefficient(s):
ar1           ar2               ar3               ar4              ar5            intercept
0.067939   0.155447   0.067962   0.026540   0.070215   0.000086
log likelihood:       5399.57
AIC Criterion:        -10785.14

Call:
armaFit(formula = ~arima(5, 1, 0), data = x)

Model:
ARIMA(5,1,0) with method: CSS-ML

Coefficient(s):
ar1          ar2            ar3           ar4          ar5
0.07284  0.15856  0.07088  0.03291  0.07734
log likelihood:       5168.35
AIC Criterion:        -10324.71

To the second answer. If the I in ARIMA just tells you how many times you need to difference, why the function armaFit makes these differenciation for you? And if it does, why it cannot be included in garchFit function in the same way? Or maybe one strange idea came to my head, what about making the ARMA model on data to get residuals, then make another ARMA model on these residuals, make 2 forecasting - forecasting of series and forecasting of errors from ARMA model and then just add these 2 results together?

I believe there is a mistake in your formula. I try it on a few data, which I diff(log) and then try to transformed it back, but it didnt work. Results was totally different. I use this formula instead, which works perfectly for me, but maybe the problem is on my side :))) :

for (t in 1:n-1)
x[t+1]]<-exp(ret[t]+ log(x[t]))

But for the prediction, I dont have the original price time series (xt) so how can I make the reverse transformation for that particular period? I mean in the case of ex-ante predictions, with the ex-post, it is OK.

And the last thing, we have learnt that only AR part needs to be stationary. So the coefficients of ARMA model for example (0,2) can be estimated on non-stationary time series.

Jan Babel

______________________________________________________________
> Komu: <babel at centrum.sk>, "R-SIG-Finance" <r-sig-finance at stat.math.ethz.ch>
> Datum: 12.04.2009 15:33
> Předmět: RE: [R-SIG-Finance] ARIMA,GARCH and differences
>
>Taking logs is not a linear transformation. Generally, values that have
been very large or very small before the transformation become more
similar after the transformation. The difference of the logs therefore
tends to be smaller than the simple difference. The simple difference
furthermore is not a percentage but an absolute difference so that scaling
matters. From what I know, most people would want to take ret and not ret1
since ret can be directly interpreted as continuous returns.
>
>For your second question: From what I know, in an ARIMA model the
coefficients are not estimated on the nonstationary series but always on
the stationary series so that the I in ARIMA just tells you how many times
you had to differences the series to estimate the ARMA model. I dont know
if I understand you correctly but the reverse transformation is easy to
do: After estimating the model on the stationary returns (ret) you
generate the forecast of the returns and then transform this back to
prices using
>
>x_t = exp[ret_t + x_t-1]
>
>where x_0 is the is the beginning of your price series. I use a loop for
the transformation although there is probably a more elegant vector
function for that.
>
>
>-----Original Message-----
>From: r-sig-finance-bounces at stat.math.ethz.ch on behalf of
babel at centrum.sk
>Sent: Sun 4/12/2009 12:46 AM
>To: R-SIG-Finance
>Subject: [R-SIG-Finance] ARIMA,GARCH and differences
>
>Dear friends
>
>This is maybe trivial question for you, but why I get different results
when I use return series and the original price series with 1.differences
speciefied in ARIMA model? For example
>x<-priceSeries
>ret<-diff(log(x))
>ret1<-diff(x)
>
>fit_arma = armaFit(~ arma(5,0), data = ret)
>fit_arma1 = armaFit(~ arma(5,0), data = ret1)
>fit_arima = armaFit(~ arima(5,1,0), data = x)
>
>The first model has the smallest AIC. AR parameters are very similar but
not the same.
>
>The second question is more critical for me. Is there any possibility to
enlarge  the package fGarch for the joint estiamtion of ARIMA+GARCH?
Because in mean equation you can specifie only AR, MA, or ARMA model, not
ARIMA. This extension I consider very important, because using the
differences biult-in model,( I mean ARIMA) simplify the entire fitting and
forecasting proces. If you use ARIMA, you dont need to make reverse
scale). For evauation part it is not a big problem but for prediction part
I find it difficult, therefore, I will prefer to use ARIMA so the garchFit
function would produce the prediction and make the reverse transformation
for me. (like the armaFit does)
>
>Thank you very much for any suggestion you provide
>
>_______________________________________________
>R-SIG-Finance at stat.math.ethz.ch mailing list
>https://stat.ethz.ch/mailman/listinfo/r-sig-finance
>-- Subscriber-posting only.
>-- If you want to post, subscribe first.
>
>
>
>EBS European Business School gemeinnuetzige GmbH - Sitz der Gesellschaft: