Granular (Upper and lower) Sensitivity

[jonasViehweger]

For some inputs (TCW, NDMI) and some types of seasonal trajectories, the direction of the residuals which signal a true break and not time-series artifacts like seasonality or missed clouds/snow is uni-directional.

For example with NDMI, true disturbances will almost always signal through sustained lower NDMI values.
Another case is residual distributions. Even though residual distributions will be centered around a mean of 0, the distribution of residuals might be quite lopsided. My guess is that this is one of the reasons why IQR performs so well out of the box; it has some amount of granular upper and lower sensitivity built in through its statistical limits.

I carried out a hyperparameter tuning test for IQR using the jr-digital/DISFOR dataset, which showed that the optimal parameters for this test set using NDMI as input and optimizing for F1 score, results in lopsided sensitivities:

{'sensitivity_lower': 2.0050327696845214,
 'sensitivity_upper': 4.259944838204534,
 'trend': False,
 'harmonic_order': 2,
 'boundary': 7,
 'L': 5.237738192464925}

compared to the optimal parameters with just a single sensitivity value for both positive and negative deviations

{'sensitivity': 2.6408646132552693,
 'trend': False,
 'harmonic_order': 4,
 'boundary': 8,
 'L': 18.57921286921607}

Being able to set sensitivities individually yields an improvement in F1 of around 3% (0.762455 with separate sensitivities vs. 0.731636).

I've implemented this additional behaviour for IQR in this pull request. There's a few things left to clear up before doing this same treatment for every class:

• Should the lower sensitivity be defined as (-lower, +upper) or how it is right now as (lower, upper) and the directionality is handled in the sensitivity calculation logic in _update_process().
• Does the docstring describe the behavirous clearly?
• Right now the Integration test is failing due to the export/import not supporting a tuple for sensitivity. I would fix that once the sensitivity is implemented for all classes.

Any thoughts @loicdtx?