I have tried forecasting in R using ets(). I let ets choose the best model for my data. The problem is i observed that eventhough the data shows an increasing trend and exhibits seasonality, ets is giving MNN model while MAM gave best results(i have chosen MAM after seeing the graph of the time series). ets selects a model based on low AIC,right? why is it selecting MNN when MAM is giving relatively low AIC value.So kindly list a procedure to forecast future values for a timeseries whose seasonality, trend are not known before hand i.e. automation of the forecast procedure.
New Edit: my data is given below:
date,value 01/08/2012,262830 01/09/2012,4849602 01/10/2012,6341298 01/11/2012,6814589 01/12/2012,9494411 01/01/2013,10559931 01/02/2013,12113638 01/03/2013,15668512 01/04/2013,933441 01/05/2013,2701218 01/06/2013,4332092 01/07/2013,7537763 01/08/2013,8485541 01/09/2013,10171206 01/10/2013,11501464 01/11/2013,11464229 01/12/2013,16046044 01/01/2014,16881837 01/02/2014,17942038 01/03/2014,22527927 01/04/2014,944640 01/05/2014,3246315 01/06/2014,5796971 01/07/2014,8759231
I used frequency=12 in ts object creation. i used na.approx to interpolate values for missing dates if any. Then i used ets with model="ZZZ" and damped=NULL. ets has chosen MNN model but the data has increasing trend and also exhibits seasonality. Shouldn't it choose MAM? Here is the graph of input and outptut
data in orange is given data
adding the code here: ('modval' is obtained after interpolation is applied to 'value' in case of missing dates)
myts<-ts(modval,frequency=12) fit<-ets(myts,model="ZZZ",damped=NULL) result<-forecast(fit,h=12,level=95) resultframe<-as.data.frame(result) pointforecasts<-resultframe[,1] lowerboundofPI<-testframe[,2] upperboundofPI<-testframe[,3]