Hi all,


I am in the process of developing an automated forecasting system for univariate time series and I would like to implement AIC comparisons of models in levels and first differences akin to Harvey (1980). His proposed method works by artifically generating a likelihood function for lets say the differenced series if the original model was in levels (or vice versa). The goal here is to have comparable likelihood functions and implicitly comparable AIC's.

The main model classes i will be working with are arima and exponential smoothing models (atleast all the additive models which are estimated via MLE in sspace as opposed to Stata's own smoothing commands). In order to derive the Likelihood function of the complement series ( first difference -> levels or levels-> first difference) it would be really convenient if i could directly grab the state space matrices from the arima command (which to the extend of my knowledge doesnt offer the matrices in e()). if i have those, i should be able to derive the likelihood function of the complement series hopefully easy enough.

I took a look at the arima.ado file but lets say theyre a bit above my coding level. Is there an easy way to get the matrices directly from the commands? If not i suppose i could set them up myself.

Thank you in advance,

Jan


Link for Harvey (1980):

https://www.jstor.org/stable/pdf/252...19095a2a45f329