I'm trying to compare the perfomance of ANN vs ARIMA and Holt-Winters for time series data, but i'm getting a huge error when using ANN to make predictions. I'm using Levenberg-Marquardt Algorithm to train the ANN and windowing technique(e.g for window of n the first n points will be my inputs and the point n+1 will be my prediction/target). I want to know how i can improve the results or forecasting accuracy of the ANN.
-> One last question: Can i please have explanation of how the
Diebold-Mariano test can be used to compare these forecasting methods(ANN,ARIMA and Holt-Winters)
This book is very good, have a look at the ANN section:
https://otexts.org/fpp2/nnetar.html
To get better results from ANNS you have to
1- Increase the amount of training data set
2- Make sure that your data set include different patterns for differnt cases
3- IF ANNS still did not give a good results then you have to one or some or all of the following
a. Increase the number of neurons in the inermadiate layer
b.Use another training algorithm
c.Use another type of ANNS
d. Use another suitable transfer function
e. Change the structure of the inermadiate layer
i feel Mazhar Hefnawi has pointed out rightly on improvement of ANN accuracy for time series.
this link is useful
https://www.sciencedirect.com/science/article/.../S009830041630060...
-
regards
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining