Fix Theta model NaN predictions

[hrzn]

Currently the Theta model sometimes predicts NaN values. This probably happens when self.season_period is equal to 0, because at some point we divide by it to get the forecast.

Another reason is that the alpha coeff of the SES is equal to 0. One trick to resolve that, is to fix the inital_level during the fit of the SES

[madtoinou]

Hi,
After inspecting the code I noticed that, the initial_level of the SES is set to 0.2 and the "reseasonalization" (where the division by self.season_period occurs) happens only if self.season_period is greater than 1. Even if the user "intentionally" sets it to 0, the "reseasonalization" does not take place thanks to the self.is_seasonal attribute.

This make me believe that this issue was addressed in another PR and could probably be closed.

[EDIT]: Found the commit fixing the issue and the corresponding PR #123.