0
$\begingroup$

Following problem. I have a daily time series from 2017 to 2021. I want to forecast the time series for each product type and different categories. I have high autocorrelation for nearly 365 lags. Any ideas for forecasting methods? Machine learning seems to be right. But as the only input is past data (no exogenous variables) Support Vector regression does not seem to be optimal.

enter image description here

Here some examples Anyone any ideas where to look for forecasting models for such type of data?

Many thanks in advance

$\endgroup$
2
  • $\begingroup$ These look quite funky, what is the process creating these? $\endgroup$
    – Tylerr
    Aug 4, 2021 at 12:44
  • $\begingroup$ These are booking data for transmission capacities for energy. $\endgroup$
    – ykn frno
    Aug 5, 2021 at 12:32

1 Answer 1

0
$\begingroup$

I dont know anything about machine learning, except that it is a popular topic nowadays. You seem eager to apply machine learning techniques to this forecasting problem, but I think some good old fashioned econometric time series analysis is suited for this problem. Your question is about what method to use, but a method as I understand it should be seen in light of some model. So I dont think it is possible to suggest a method when there is no model. I would say that given you only have historical data on endogenous variable(s) with strong autocorrelation I would think applying (V)ARMA(p,q) type models is the way to go. Maximum likelihood estimation is typically the method to determine the model parameters given your data.

$\endgroup$
3
  • $\begingroup$ What you suggest IS machine learning, it's simply using a machine to learn parameters although a lot of more traditional stats practitioners reserve it for non-parametric methods like trees or neural nets. Time series especially has been quite resistant to the term, see this excellent answer: stats.stackexchange.com/questions/160382/… $\endgroup$
    – Tylerr
    Aug 4, 2021 at 12:41
  • $\begingroup$ I tried to model it with arima. But auto.arima in r run on error. Also acf showed high autocorrelation for more than 300 lags. The time series is not stationary and i am not sure how to make it so. Differencing might not be possible here $\endgroup$
    – ykn frno
    Aug 5, 2021 at 12:34
  • $\begingroup$ A colleague suggested the holt winters approach. But as I explained arima might not be working here due to the data being non stationary. $\endgroup$
    – ykn frno
    Aug 5, 2021 at 13:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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