#################################
### R commands used in Lecture 7:
#################################
> require(qunatmod)
> getSymbols("SPY",from="2003-01-03",to="2017-04-30")
> head(SPY)
> rt = diff(log(as.numeric(SPY[,6])))
> source("mvwindow.R")
> mvwindow(rt,63)
> mvwindow(rt,126)

### Comamnds for the GM and SP example are in the lecture note.

### R package for the genreal Threshold AR models is TSA with command tar.
### R package for Markov switching models is MSwM with command msmFit.

> library(nnet)   % Load neural networks package
> da=read.table("m-ibmln2699.txt") %%% Monthly  IBM  log returns
> dim(da)
[1] 888   1
#### Use the first 864 data in training and the last 24 data points for forecasting
> x=da[,1]
> y=x[4:864]
> ibm.x=cbind(x[3:863],x[2:862],x[1:861])
> ibm.nn=nnet(ibm.x,y,size=2,linout=T,skip=T,maxit=10000)
> summary(ibm.nn)
a 3-2-1 network with 14 weights
options were - skip-layer connections  linear output units 
  b->h1  i1->h1  i2->h1  i3->h1    # Coefficients for the h1 (First node of Hidden layer)
-368.93  568.82 -295.27 -885.04 
  b->h2  i1->h2  i2->h2  i3->h2    # coefficients for the h2.
  91.71   40.63 -181.32 -218.14 
   b->o   h1->o   h2->o   i1->o   i2->o   i3->o  # Coefficients for output node.
   1.01   -3.17    3.25    0.16    0.07   -0.03 
> sse=sum((y-predict(ibm.nn,ibm.x)^2))  # Compute sum of squares of residuals.
> sse
[1] -1483.777
> ibm.p=cbind(x[864:887],x[863:886],x[862:885])
> yh=predict(ibm.nn,ibm.p)   # Prediction
> yo=x[865:888]
> ssfe=sum((yo-yh)^2)  # Compute sum of squares of forecast errors.
> ssfe
[1] 2125.665
> 
**** You can run the backtest using neural networks ("backnnet")
> source("backnnet.R")
> m1=backnnet(ibm.x,y,840) # here 840 is the starting forecast origin.