[ML] Large outliers can impair our ability to correctly identify and model periodicity

[tveasey]

Currently, we don't have explicit handling of outliers when 1) testing for periodic patterns in the data and 2) when initialising new seasonal components. This can sometimes slowdown or prevent us detecting the correct periodic components present in a signal. We had a support case raised where this was the underlying reason for us failing to detect an outage with sufficient severity.

[tveasey]

I have a reasonable implementation for this. The strategy is relatively simple: we iteratively refine the definition of outliers with respect to the period under test and reweight those samples when computing test statistics such as variance reduction and cyclic autocorrelation. I use the same iterative reweighting scheme to initialise the seasonal components. As a result, we get faster detection of the correct seasonal components in signals which are distorted by anomalous features and more accurate initial modelling.

For example,
Before:

After:

Similarly,
Before:

After:

[tveasey]

Closing as I don't have any further plans than what was implemented in #90 at present.