From my reading, many Time Series Machine Learning technique treat the whole time series as single data set and try to learn from that.
I am how to learn from a collection of series rather than one.
For example, considering the series
2020-01-01 ---- 2020-01-02 ---- ...... ---- 2020-06-30
My hypothesis is that there are patterns in each shorter periods; the information so far back in the past is not impactful and just causes noise.
Instead of learning this as a single series, I would like to learn the most common pattern, if any, for shorter chunk of periods. The training data will be split into multiple series like this
2020-01-01 ---- ..... ---- 2020-01-31 2020-02-01 ---- ..... ---- 2020-02-29 .... 2020-06-01 ---- ..... ---- 2020-06-30
Please note that this is just example. I don't necessarily look at 1-month window. I would like to explore different window as well, and identify what is the best duration for this window.
The idea remains the same : How do we learn the common pattern from many small chunks rather than trying to fit one long series?