How to use roll_time_series function properly?

[Mahyar24]

Hi,

I have a time-series dataset with minute-by-minute data. I would like to get the last 4 hours of data to predict a classification label for the next 1-hour. To achieve this, I created IDs using pd.Grouper(freq="4H") to group every 4-hours together. However, when I don't use the roll_time_series function, the resulting tabularized dataset will look like this:

2023-01-01 00:00:00 -> 2023-01-01 04:00:00
2023-01-01 04:00:00 -> 2023-01-01 08:00:00

Unfortunately, in this format, I am losing lots of data. What I actually want is:

2023-01-01 00:00:00 -> 2023-01-01 04:00:00
2023-01-01 01:00:00 -> 2023-01-01 05:00:00

I'm not sure how to create it accordingly with the roll_time_series function. Perhaps it should involve using parameters such as max_timeshift and min_timeshift, but I have no clue what values these parameters should take.

P.S. Here's an example of my data structure:

date	feature_1	feature_2	id
2023-01-21 00:00:00	-0.260733	-9999	20230121000000
2023-01-21 00:01:00	-0.263591	-9999	20230121000000

[nils-braun]

Have you looked into the documentation for the function? https://tsfresh.readthedocs.io/en/latest/api/tsfresh.utilities.html#tsfresh.utilities.dataframe_functions.roll_time_series
There is also a dedicated document text for rolling: https://tsfresh.readthedocs.io/en/latest/text/forecasting.html#forecasting-label

There you will find what these arguments mean. For your use-case, you can set the max_timeshift to 4*60 (maximum length of each resulting time series) and the rolling_direction to 60 (the size of each step). Have you actually tried the function already and are facing any errors with it?

[Mahyar24]

Yep, I've read it and appreciate it. IMHO, however, it isn't clear enough (at least for me!). I was tinkering with max_timeshift and min_timeshift, but they didn't produce the desired outcome. Additionally, I ignored rolling_direction because, based on the variable name, I thought it was irrelevant to my purpose. Sorry about that and thanks.

[nils-braun]

No worries - it's good you asked here! Do you have a suggestion on how to improve it? Feel free to do a PR or tell us about it :) Happy to incorporate your feedback!

[Mahyar24]

It's hard for me to wrap my mind around the source code, and eventually, I had to write a helper function for myself. However, I highly recommend renaming rolling_direction to something like steps and using annotations corresponding to variable names on the images in the docs. It makes it so much easier to understand.