2

If you have multivariate (ie n time series) and you want to predict one of these time series using data from every column, you could use a VAR model (Vector Autoregression).

  1. What I dont understand is why I havent seen VAR model with moving averages taken into account, like with univariate ARMA models. Do they exist or why dont they?

  2. What are the alternatives to this for predicting one time series, using data from n time series? See 1.2 in here https://www.analyticsvidhya.com/blog/2018/09/multivariate-time-series-guide-forecasting-modeling-python-codes/

Trajan
  • 369
  • 4
  • 17
  • VARMA models have been known for a long time, but VAR have been more popular for their ease of estimation. Estimation of VARMA in high dimensions is quite prohibitive computationally. But see the `bigtime` package as mentioned [here](https://stats.stackexchange.com/questions/152202/regularization-for-arima-models/295618#295618) for a recent advance in that problem. – Richard Hardy Oct 11 '20 at 18:26
  • You might also want to learn about [empirical dynamic modeling](https://link.springer.com/article/10.1007/s11284-017-1469-9)—including [simplex projection](https://rmgsc.cr.usgs.gov/outgoing/threshold_articles/Sugihara_May1990.pdf), [S-maps](https://royalsocietypublishing.org/doi/abs/10.1098/rsta.1994.0106), [convergent cross-mapping](http://www.uvm.edu/~cdanfort/csc-reading-group/sugihara-causality-science-2012.pdf), and related methods—which provide a *very* different approach to time series prediction from VAR. – Alexis Oct 11 '20 at 18:36
  • @Alexis i will look into those but i was hoping for things more simple. Are there issues regarding using standard ml algorithm with the dependent variable lagged 1 time step into the future? – Trajan Oct 16 '20 at 10:57
  • So you are looking for alternatives, but are not actually interested in alternatives? Can you clarify what the second part of your question is asking? – Alexis Oct 17 '20 at 17:39

0 Answers0