Mos Rah By ML, if you are specifically looking for Neural Network-based approach, perhaps LSTM (Long Short Term Memory) is the best example for Time-series applications.

VAR is acronym for vector autoregressive model. An autoregressive model is a linear combination of lagged elements of the same variable and items of a white noise process. The estimate of the white noise process gives the shocks.

I think you can use recurrent neural network (RNN) architecture, precisely :Long short-term memory (LSTM)

more information is in this paper :

Article A novel Hybrid Bidirectional Unidirectional LSTM Network for...

Ette Etuk Ashutosh Karna Anouar Ben Khalifa

Thanks for your help.

So In brief: I should use LSTM ,And estimate the white noise process.

I think I must do this process:

A. Calculate fluctuations of my variables.

B. Run LSTM for estimate the white noise.

Am I correct? After section`B` what should I do?

Good luck