*
* Neural network example from pp 180
*
open data ch04_ibm_m(26-99).txt
calendar(m) 1926
data(format=free,org=columns) 1926:1 1999:12 ibmln
*
* Benchmark - use mean of data through 1997:12
*
stats ibmln * 1997:12
set benchfore 1998:1 1999:12 = %mean
*
@uforeerrors ibmln benchfore
*
* AR(1) model
*
linreg(define=ar1eq) ibmln * 1997:12
# constant ibmln{1}
*
* Out-of-sample static (one step) forecasts for the AR(1)
*
uforecast(equation=ar1eq,static) ar1fore 1998:1 1999:12
@uforeerrors ibmln ar1fore
*
* Train the neural net. This sets an R**2 target of .10. This is probably
* unachievable, but with only two hidden layers, it's unlikely that we
* have to fear overfitting.
*
nnlearn(rsquared=.10,iters=100000,hidden=2,direct,save=nnmodel) * 1997:12
# ibmln{1 2 3}
# ibmln
*
* Out-of-sample static forecasts for the neural net. Note that these are the
* only types of forecasts that the neural net can produce (easily). To do
* multiple step forecasts, you would need to loop over a series of one step
* forecasts and feed the forecasts back into the series.
*
nntest 1998:1 1999:12 nnmodel
# ibmln{1 2 3}
# nnfore
@uforeerrors ibmln nnfore