[ML] Forecasting trend is unexpected given data

[valeriy42]

We observed that forecasting can exhibit unexpected and unintuitive trends given data with interruptions:

or temporary decrease:

I was able to reproduce the behavior using synthetic data:

May be related to #513

[valeriy42]

The problem is that for a certain unfortunate window of time, the data may look like it contains an upward trend from the model's perspective:

I am looking at how to fix this behavior without affecting the generally correct working trend recognition.