Extracting features of several timeseries

[arresejo]

Hello,

First of all, congratulations for the amazing work you did !
I'm quite new to the word of time series, sorry in advance if the questions seem too obvious :)

I'm currently trying to use tsfresh to extract useful features of several currency pairs time series (Forex) to perform some clustering in order to find some correlation between pairs.

So first, I retrieve all the raw data for several currency pairs and append everything to build the dataframe in the format expected by tsfresh:

My questions:

1. Is it the good approach to pass directly this big dataframe to extract_features or should I rather use a rolling window approach ?
2. As I'm working on historical data, is it possible to perform the selection of relevant features ?
3. Do you recommend to perform some scaling (e.g. min-max) before calling extract_features ?

Thanks in advance,
Jorge

[kempa-liehr]

Hi Jorge,
Thanks for your encouraging feedback.

1. Passing a time-series of this length to tsfresh would only work for the minimal feature set due to constraints of computational complexity. So, I suggest to use a rolling window.
2. Yes, that is possible, have a look at the following article, which describes the idea of feature selection for clustering: https://ieeexplore.ieee.org/document/9444391
3. Yes, you should consider scaling.

Cheers,
Andreas

[arresejo]

Thanks @kempa-liehr, really appreciate !

Cheers,
Jorge