Closed config-i1 closed 3 years ago
Use principles in https://openforecast.org/adam/ets-arima-1.html to kick out the unnecessary orders.
Part of this is done in #161. Updated the original issue description.
(1) and (2) are now done in 1969962c1e33027c1cbb88b0ae5e5978624786f6. The initial states are still far from ideal, but at least it works faster than before and it is more accurate than before.
Another unresolved issue - selection between seasonal and non-seasonal ARIMA on difficult data with initial="optimal"
. Currently, the seasonal ones are selected only on very well pronounced seasonal data. The selection works fine in case of backcasting.
Looks like the majority of issues are resolved now. The initialisation of ARIMA is still not perfect, but works.