Auto forecast: resistant to model errors

[dareid]

Using the automatic forecasting option for a small dataset always results in an error from the TripleExponentialSmoothing model which states it needs 2 years of data, which is fair enough.

However, I don't think this should kill the automatic forecasting. Maybe errors from models could be caught, and only the models which successfully built could be assessed?

[nlaptev]

Once again dareid, nice catch. Please submit a pull request for this.