Can we really forecast using Time series Neural Network

Hi everyone,

It's been a while that I'm looking how I can forecast a time series data using NARX.

I think the NARX is the most suitable model for my data so far but I want to forecast in the future and it seems to be not possible. Although, I've been looking at the different ansewrs and there are some very interesting but it's not working.

I have 300 time step data and I want to forecast 120 step in the future. Does anybody have something clear to show me. Or should I model in a different way and not use the Neural Network.

Thank you

Best Reagards,

Nadia

[Akash Menon]

https://towardsdatascience.com/how-not-to-use-machine-learning-for-time-series-forecasting-avoiding-the-pitfalls-19f9d7adf424

Check that out. It depends on your type of precidtion you want to make. Using past data to predict future data is a stochastic approach & may/may not be applicable in your case.

NARX is dependent on an eXogenous (external) input. So to accompany your 300 time steps there should be some input parameter that the output depends on. So to make those additional 120 step prediction you will need the future inputs for the next 120 steps too.

300 timsteps seems a bit small to train a neural network but if you have more data with corresponding inputs, then you can train the network well & deploy it for multi-step prediction in the closed loop mode.

See: https://nl.mathworks.com/help/deeplearning/ug/multistep-neural-network-prediction.html

Hope that helps!