Made a forecast with a DS serie Rstudio

[Paulo13]

Hi everyone,

I have some troubles with my university homework.
I'm actually learning rstudio.
Teacher give us a big homework.

I will try to explain (i'm french and english is hard for me ^^) :

I downloaded a csv file with the interest rate of canada from 1955 to 2021.
--> I made a lot of test : DF , ZA and LS test. So i found a DS serie with structural change with the LS Test.

I differentiated and stationnarised the DS serie to made forecast on it.

Then i'm made ACF/PACF to see if there is autocorrelation is my stastionnarised serie i found a little autocorrelation at lag 6.
But when i made ljungbox with lag = 10, i found p-value = 0.4 who correspond to nul hypothesis of autocorrelation...
When i made eacf function i found ARMA(0,0) ...

I'm lost ^^

Then i try to made my model with the best bic , i found

Arima(dCanada, order = c(1,6,0),include.mean=FALSE) with a BIC of 483 (ar(1) best then ma(1))
And Arima(dCanada, order = c(0,0,0),include.mean=FALSE) with a BIC of 180

So i conclude the ar(1) is the better model ?

The second problem is this :
-LjungBox test on centered and reduced residus has autocorrelation (p-value<0.05)
-ArchTest fund cluster of volatility ...

So which model do i choose ?
My forecast isn't possible?

Thanks,
Paul

[phil_hummel]

So you have interest rates vs time and the question is can you accurately forecast and for how long... Did you split your history into training and testing periods? How good was your model vs reality?

[Paulo13]

Hi phil thanks for your answer.

I send you my model

fitdCanada = Arima(dCanada, order = c(3,0,3),include.mean=FALSE)
coeftest(fitdCanada)
BIC(fitdCanada)

The BIC is equal to 177, all coefficients and all hypothesis are OK.

But... Eacf give me the best model for ARMA(0,0) and when i made a c(0,0,0) in Arima, I found a bic of 480.
But in this case, my model don't validate white noise hypothesis and my model isn't good.

Thanks