2
$\begingroup$

I would like to compare one-step ahead forecasts on a given time series for ARIMA and UCM (using KFAS library). I have split my time series in train and validation, that I will use to understand which model performs better.

After reading the anwswer to this post: Difference between first one-step ahead forecast and first forecast from fitted model, I understood the following explanation for Arima:

So fitted(fit) gives one-step forecasts of observations 1, 2, ... It is possible to produce a "forecast" for observation 1 as a forecast is simply the expected value of that observation given the model and any preceding history.

fitted(refit) gives one-step forecasts of observations 401, 402, .... So it uses the model estimated on observations 1...400, but it uses the data from time 401...500.

What I don't understand is how to make one-step forecasts on observations 401, 402, .. using data from time 1.. 400 and (possibly?) the calculated forecasts up to previous time. I want to pretend not to know data 401..500 and compare the one-step forecasts with them.

Thank you for any help on this.
Francesca

$\endgroup$
0
$\begingroup$

(hi, i cant comment so

1. arima (1,0,0) = ar (1) need 1 previous observation ar (2) need 2 previous observations

2. arima do calculate with training data value of coeficients (for ar (1) its like: intercept and ar1

you can calculate coeficients arima(data[1:400] or you can also arima(data[1:472]

(1) so you always need previous (current data) to make 1 step prediction

(2) you can use the same values of coeficients or you can update them as new data are comming

$\endgroup$
2
  • $\begingroup$ Thank you for your help. Practically speaking, should I loop 100 times, applying forecast using the previously trained cofficients , h=1, and the train timeseries followed by the already forecasted observations? I tried this: predictions <- c(rep(NA,len_validation)) # last 100 observations temp_ts <- ts_train # first 400 observations for (i in 1:len_validation) { temp_mod<- Arima(temp_ts, model = mod1.5) predictions[i] <- forecast(temp_mod, h = 1)$mean temp_ts <- ts(c(temp_ts,predictions[i])) } but it seems to return the same as forecast(mod1.5, h=len_validation) $\endgroup$
    – botti23
    Feb 26 at 5:12
  • $\begingroup$ (i didnt work with re-using model..i can onlyguess: if at 1:400 , prediction for h=100 ....for example 500th data point - is 100% exactly same as model trained 1:499 for h=1.....it tells that data are "very deterministic" ... how can i guess weather for 100th day,..with same precision as waited 99 days and guess tomorrow .... but i checked rdocumentation.org/packages/forecast/versions/8.13/topics/Arima and it is possible - i think it points, that data are strong deterministic, cycled....if its for school, i personaly would mention both results and point to good data stability : ) $\endgroup$
    – user2120
    Feb 26 at 11:14

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.