forecasting error?

Hello everybody!
I have an ARIMA model for a time series. This model was obtained through an
auto.arima function. The resulting model is a ARIMA(2,1,4)(2,0,1)[12] with
drift (my time series has monthly data). Then  I perform a 12-step ahead
forecast to the cited model... so far so good... but when I look the plot of
my forecast I see that the result is really far from the behavior of my time
series... in fact, there is a considerable gar between  the last value of
the series and the first forecast. My guess is that I'm doing something
wrong. Here is what I do:

mods<-auto.arima(x[[1]],start.p=0,start.q=0,start.P=0,start.Q=0,stepwise=TRUE,stationary=FALSE)
ARIMA(2,1,4)(2,0,1)[12] with drift         # the output

Call: auto.arima(x = x[[k]], start.p = 0, start.q = 0, start.P = 0, start.Q
= 0, stationary = FALSE,      stepwise = TRUE)

Coefficients:
        ar1      ar2      ma1     ma2      ma3     ma4    sar1
     0.0639  -0.7820  -1.2103  1.2236  -0.9511  0.2357  1.0031
s.e.  0.0686   0.0582   0.1098  0.1558   0.1568  0.1007  0.0716
        sar2     sma1      drift
     -0.0711  -0.8963  -780.9456
s.e.   0.0747   0.0608   403.2112

sigma^2 estimated as 10202381:  log likelihood = -1100.61
AIC = 2206.69   AICc = 2209.23   BIC = 2236.98

 >for<-forecast(mods,h=12,newxreg=(1+length(x[[1]])):(length(x[[1]]+12)))

#forecast

and as I said before, the results dont seem to be right. In fact, when I
restrict the search of the model on the auto.arima function to stationary
models only an I perform the forecast (without the newxreg-option) the
results are very much acceptable.

ANY HELP OR COMMENTARY I VERY WELCOMED!!!!!!  thanks in advance!


Diego.

       [[alternative HTML version deleted]]