2
$\begingroup$

When we have to make a forecast, the books tell us that the main method is the autoregressive moving average model. In my opinion there is another big tool, the feed forward neural network (FFNN). So I think that we could use two main tools:

  • Autoregressive moving average
  • Feed forward neural network

Of course there must be differences, but I am not an expert. Who, having sufficient experience in these two methods, can explain to me the differences between these two methods on making predictions?

$\endgroup$
3
$\begingroup$

Feed forward neural networks, as opposed to ARMA models, attempt to capture the fundamental patterns present in your data. In other words, ARMA predicts the next step based on the moving average and its current tendency. FFNNs try to see the bigger picture and approximate a non-linear function that maps previous steps to the next step. Theoretically, FFNNs may give you better long-term predictions. However, training FFNNs is no easy task; the number of different FFNN training algorithms and corresponding parameters to be optimised is somewhat overwhelming. The major difference between the two approaches can be formulated as follows: ARMA produces linear models; FFNN produces non-linear models. If your data is simple and can be fitted to a linear model, you are better off using ARMA. If you suspect non-linear relationships between variables are present in your data, or if ARMA does not produce a model of sufficient quality, go for FFNNs.

$\endgroup$
  • $\begingroup$ Thanx, good answer, but one doubt. How can I test the forecast performance between them? $\endgroup$ – emanuele Feb 21 '14 at 12:24
  • 1
    $\begingroup$ They were. Look here. neural-forecasting-competition.com/NN3/results.htm Also, look for our software listed as "Autobox/Reilly". The most important thing not being discussed here is that you need to consider outliers and adjust for them. You might also have changes in trend/level/seasonality/parameters/variance. $\endgroup$ – Tom Reilly Feb 21 '14 at 12:57
  • 1
    $\begingroup$ @emanuele, there is more than one one to measure forecast performance. The simplest one is to calculate the generalisation error (error produced on the data that the model was not exposed to during training). $\endgroup$ – anna-earwen Feb 25 '14 at 6:52

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.