python - ARIMA out of sample prediction in statsmodels? -

i have timeseries forecasting problem using statsmodels python package address. evaluating using aic criteria, optimal model turns out quite complex, arima(27,1,8) [ haven't done exhaustive search of parameter space, seems @ minima around there]. having real trouble validating , forecasting model though, because takes long time (hours) train single model instance, doing repeated tests difficult.

in case, need minimum in order able use statsmodels in operations (assuming can model validated somehow first) mechanism incorporating new data arrives in order make next set of forecasts. able fit model on available data, pickle it, , unpickle later when next datapoint available , incorporate updated set of forecasts. @ moment have re-fit model each time new data becomes available, said takes long time.

i had @ this question address problem have arma models. arima case there added complexity of data being differenced. need able produce new forecasts of original timeseries (c.f. typ='levels' keyword in arimaresultswrapper.predict method). it's understanding statsmodels cannot @ present, components of existing functionality need use in order write myself?

edit: using transparams=true, prediction process needs able transform predictions original timeseries, additional difficulty in homebrew approach.

an arima(27,1,8) model extremely complex, in scheme of things. time series, can reasonable prediction 5 or parameters. of course depends on data , domain, i'm skeptical 27 + 8 = 35 parameters necessary.

the aic known permissive number of parameters. i'd try comparing results bic.

i'd whether data has seasonality of kind. e.g., maybe 27 of ar terms don't matter, , need lag=1, , lag=24 (for instance). might case hourly data has daily seasonality.