I am trying to understand timeseries analysis. I ran ARIMA model for same, the values i got were...
p,i,q =(7, 1, 1)
Now i created a subset from this dataset by applying certain external filters and for these same value of (p,i,q) I ran the ARIMA model.
but i got error...
The computed initial AR coefficients are not stationary
You should induce stationarity, choose a different model order, or you can
pass your own start_params.
The below mentioned data is completely fictional, mentioned just to explain what i did.
T being the timestamp and A is the value being predicted.Sample dataset:
T, A B F
1, 10, 1, 4
2, 60, 0, 13
3, 40, 2, 9
4, 20, 1, 7
I created ARIMA model and suppose the value i got were...
p=7, q=1, I=1
So i made another dataset out of it as...
A <-- IF(B>1: A=A-C, ELSE A=A)
i.e. if B is greater than 1 A is A-C, else A is as it is.
So this equation gave me new values of A which i again ran with same value of p, q, i. But I got that non-stationarity error.
Now couple of queries that i had are:
I think the way i changed the column values, it completely changes the data which clearly should not be used with previous value of p,q,i, we must again check the data and check for ARIMA model.
Secondly, I went and check out of nowhere for a couple of values of p,q,i=[(0,1,2) and (0,1,1)] and applied ARIMA. I saw that there was an accuracy of 99.5% accuracy which seems to be... well, not something i expected. What does it means when we get exact predicted values as actual values with Timeseries model. It cannot be correct I think.