tbats fixes

[jmoralez]

Makes the following bug fixes:

• Applies the transformation with the optimum lambda to the seed states when using boxcox.
• Removes the Augmented Dickey-Fuller test.
• Checks if use_trend in (True, None) to correctly generate the combinations in the auto model.
• Renames seasonal_periods to season_length to be consistent with the other models.

Also modifies the process to find the optimal number of harmonics by:

• Using an expanding mean if there aren't enough samples in the window to remove the trend (thus doesn't produce missing values). We discussed if we should use center=True and found that overall the forecasting error is smaller when using center=False.
• Subtracting the signal found for each seasonal period before looking for the next one

Adds an experiment with the M4 hourly dataset.

[review-notebook-app[bot]]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[MMenchero]

Important changes made by @jmoralez:

• Uses a new Box-Cox transformation that fixes the cases where the inverse transformation results in missing values.
• Limits the number of searches when determining the number of harmonics.
• Adds a scale parameter to the objective function, akin to what R does. With these changes, I reran the complete experiments for the M3 and M4 datasets.