Econometric vs ANN models for forecast?

Question

I hope this is an appropriate question for this forum... for me it is an obvious query since it intrigues me for a long time.

Ok, assume there are 2 distinct classes of models: econometric (AR, MA, ARIMA, ARCH, GARCH, EGARCH, TGARCH, ...) and neural networks (MLP, RBF, BPTT, TDNN, Elman, NARX, ..., I'm putting even SVM and SVR into this group).

I know it is a broad subject - depends on the market, assets, for a start... but under what conditions one is better than the other? Is there a general consensus over this? In terms of MSE, R², accuracy and so on? Is it fare to compare them? Does it make sense?

I've seen many studies doing this kind of comparison, here is an example (sorry for this being biased towards one side). But none summarizing previous conclusions on this topic.

Finally, what is your experience with both of them? Do you have other articles running this kind of test (even if not published)?

Thanks in advance, DBS.

Answer 1

They are not mutually exclusive. For example, the class you refer to as "econometric" are simply linear regression models that include as factors prior returns or residuals of the return series sometimes with weightings on the observations.

You could easily design a neural network with no hidden layers and the same inputs. So each of the econometric models are special cases of neural network models. Neural Networks offer a broader class of modeling options although for the same reason they are more difficult to train and avoid overfitting.

In the paper, the authors are not comparing neural networks as a class to the econometric models. They are comparing a very specific neural network configuration to traditional time-series models.