# "Variable time steps in Time Series analyses"

Hi all,

Great job with RM! I am having a great time learning all the different functionalities.

Checking out the time series processes, one doubt came up. Is it possible to run analyses in series with variable time steps? I couldn't find anything that would allow me to associate a time vector to my data. In Time Series, is time always an integer counter?

I really appreciate all the help.

Joao G.

Great job with RM! I am having a great time learning all the different functionalities.

Checking out the time series processes, one doubt came up. Is it possible to run analyses in series with variable time steps? I couldn't find anything that would allow me to associate a time vector to my data. In Time Series, is time always an integer counter?

I really appreciate all the help.

Joao G.

Tagged:

0

## Answers

537MavenThen a learner can find a relation between the target and these attributes.

That said, the "windowing" operator kind of assumes that your samples are equally spaced out in time.

If they are not, they are still treated as equally spaced out in time.

For example if you have series data:

x0 ... x10, window size 3, horizon 2, you get

x-2, x-1, x-0, x+2

x0, x1, x2, x4

x1, x2, x3, x5

x2, x3, x4, x6

x3, x4, x5, x7

x4, x5, x6, x8

x5, x6, x7, x9

x6, x7, x8, x10

So in this example you are always predicting 2 samples into the future.

Based on the last 3 samples.

When samples lie at different intervals, you might sometimes be predicting hours into the future, while sometimes only minutes.

This is a very weird situation.

It is therefore common to aggregate your data so that samples becomes equally spaced out in time.

For example, if you have observations roughly each minute, but a lot of observations are missing, like you have observations form minutes:

0 1 5 6 8 9 11 12 14 15 .. etc

Might be a good idea to aggregate your data in periods of 5 minutes, so you get

5, 10, 15, 20, etc, which are nicely equally spaced.