I'm not sure if I really get your point, but I think you have several options:

1. simplest way, use the time since start as predictor ( 0 ms, 10 ms, 20 ms and so on)

2. you can use the predictors of the previous and succeeding time steps as additional predictors (take care of the first and last time step, some packages can't deal with NA)

Maybe not in your case, but for longer time series:

3. if you have longer time series, use the time of the day (0 to 24) as predictor

4. if you have long time series, use tday of the year (0 to 366) as predictor

Hi Johannes

These guys used treeBagger for their time series predictions - maybe some of their ideas will work for you:

Conference Paper Multi-step Ahead Time Series Prediction via Bagging Trees Ba...

Time series can be characterized by many properties:

1) time series mean, median, sd, length, min/max, etc. (these are obtained straightforwardly)

2) the presence of monotonic trend in location. This can, e.g., be based on coefficients obtained using the Theil-Sen estimator

3) the presence of monotonic trend in variance. The same, but based on squared/absolute values

4) the presence of cyclical trends in location (see Runs test)

5) the presence of cyclical trends in variance

6) seasonality

7) structure, the presence of AR/MA terms (can be based on ARIMA modelling)

etc..

One problem with time series models to produce time series characterizations is that such models usually require stable variance. To get stable variance time series you can use this guide:

https://xdplot.blogspot.com/p/data-transformations-guidelines.html

So, perhaps, for velocity you need to log-transform time series prior to obtaining the above params.