Forecasting Econometric Methods-Neural Network

[simoni sour]

Hi everyone!

I would like to conduct a research concerning the energy demand forecasting. Unfortunately, my data is stationary Dickey Fuller test and Phillips Perron also verified the stationarity and also the graphical repersentation of the data. So, I cannot employ an ARIMA model or a State-Space. Do you have any idea about different forecasting methods with stationary time-series? Typically, I've run an ARMA model what else do you think I can do? Thank you for your advice. The scope of my paper is the comparison of different econometric approaches, but I cannot follow the Box-Jenkins methodology as long as my time series is I(0). Additionally, I would like to ask you if there is a code for Artificial Neural Networks. Thank you for your help!

[Anton Belgrave]

Hi
As I understand it, you can still use the state space framework except that you would drop the trend. Stata has a variant of the state space called unobserved components that is fairly approachable. Variants of the traditional econometrics methodology would also be valid if your series are I(0). Unfortunately, I am unfamiliar with neural nets although there are user written packages for this in Stata’s framework. A better source may be found in other languages which seem stronger in this area. Try searching for long short term memory models. However, before going there, you should also note that while these frameworks may be good for forecasting you may find that their results are black boxes, that is, difficult to explain.
Regards
Anton Belgrave

[Nick Cox]

Anton Belgrave gave good advice.

Box-Jenkins includes ARMA within ARIMA. Indeed it includes AR and MA too. Being approximately stationary is not a problem (although it's hard to believe for energy demand).

[simoni sour]

Hello! Thank you for your suggestions!! use daily data for Germany's load from ENTSO-E transparency. Indeed, it's hard to believe that energy demand is stationary. Here is the D.Fuller results at level.
dfuller lnMaxload1

Dickey-Fuller test for unit root Number of obs = 398

Interpolated Dickey-Fuller --------
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value

Z(t) -8.239 -3.448 -2.874 -2.570

MacKinnon approximate p-value for Z(t) = 0.0000


I also upload the graph. I have a question: In order to fit ARIMA my data must be integrated, if my data is I(0) then I have to perform an AR, MA and ARMA model. If I perform the forecast how can I get the MAPE, MAE, RMSE? Thank you for your help once again!!