# Neural network for multivariate time series

I am currently analysing multivariate time series and have worked on VAR models in R. I need to know if there is any way to analyse it using neural networks in R.

PS: I am aware of the nnet package and tsDyn package, but I guess they are applicable for univariate time series only.

Perhaps you can put different time series with lags on columns and predict each individual time series by its own lags and lags of other variables.

VAR models are consistently estimated by OLS equation by equation if each equation contain same variables (namely variables own lags and lags of other variables). You could use this principle.

Regards,

Analyst

• Thats Right Analyst. In fact I've been doing that. But then everytime I have to decide how many lags of each variable to use in the model. I have to make model to predict demand for 100+ products, I was looking for some kind of automation. – NG_21 Nov 28 '13 at 9:18
• For the record: VAR models are consistently estimated by OLS regardless of whether all equations contain the same set of variables. It is the relative efficiency of the OLS estimator as compared with GLS estimator that depends on whether all equations contain the same set of variables; OLS is less efficient than GLS when the variables are not the same in all equations. – Richard Hardy Feb 11 '15 at 13:43

I do not know much about R, but if there are methods to analyse univariate time series, probably the first step will be to make the multivariate data univariate. I assume you have a tuple containing different types (e.g. a mixture of bool, emueration or real valued data) for each time step.

You could e.g. take an Restricted Boltzman Machine like algorithm and extract the features of thoses tuples into a univariate real (or boolean) valued vector. Then you feed this univariate time series into the R routienes.

This paper describes a way to feed multivariate int a RBM. I found it quite nice! :)

Be aware, that rebuilding their work (in R?) will likely cost a significant amount of time!

• @gung Sorry for beeing that brief in the first place. I changed the link to directly point to the specific paper on arxiv. Explaining their work here is likely an overkill. But i consider this paper very helpfull for the related task. At least it might point into the right direction :-) – Marti Nito Feb 25 '15 at 15:15