Disclaimer: Not from math background

I want to use neural networks for forecasting a time series data. I am reading/watching basics about ACF and PACF. While reading about those functions I came to understand that PACF is an important function to find different time series models for example AR(1), AR(2) etc which tells us which lagged time series is important for forecasting.

Image courtesy: http://stats.stackexchange.com/questions/106038/estimate-arma-coefficients-through-acf-and-pacf-inspection

In the above PACF image time series lagged at 1,2,12,24 are more prominent and above confidence level. Can we use these time series lagged data Y(t-1),Y(t-2),Y(t-12), and Y(t-24) as the feature vectors for neural network for training to forecast Y(t+n)?


Yes, this is usually the case when creating autoregressive (AR) models, as you can see in this question: Estimate ARMA coefficients through ACF and PACF inspection

However, ACF and PACF compute linear correlations, while neural networks are nonlinear, so the applicability of those techniques may not be obvious. This paper shows that they are indeed useful for feature selection on neural networks.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.