###
###
###  Neural Network Demo 3 - Smarket Data
###
###
####################################################
###                        Also see ML-04Lab2.R






###-------------------------------------------------
###--- 1. Preliminary
library (ISLR)
names(Smarket )

    # [1] "Year"   "Lag1"    "Lag2"   "Lag3"   "Lag4"
    # [6] "Lag5"   "Volume"  "Today"  "Direction"

dim(Smarket )
# [1] 1250 9

summary(Smarket)
pairs(Smarket)

cor(Smarket)             # gives error because the 1st column is not numeric
cor(Smarket[,-c(1,9)])   # 1st col is Year, 9th col is Up/Down response
attach(Smarket)







##--- Scale data for neural network
Means = c(0, apply(Smarket[,2:8], 2, mean))
SDs   = c(1, apply(Smarket[,2:8], 2, sd))
Smarket.sc = as.data.frame((Smarket[,1:8]-Means)/SDs)
Smarket.sc = cbind(Smarket.sc, "Direction"=Smarket[ ,9])


##--- Separate into Training Set and Testing set ---

ix.train = (Year<2005)               # Train = 2001 to 2004
                                     # Test  = 2005

Smarket.tR  <- Smarket.sc[ix.train,  ]    # Training Set
Direc.tR    <- Direction[ix.train]    # Training Response

Smarket.tS  <- Smarket.sc[!ix.train, ]   # Test Set
Direc.tS    <- Direction[!ix.train]   # Test Response

dim(Smarket.tR)
dim(Smarket.tS)
#[1] 252 9




###-------- Fit Neural Network
library(neuralnet)          # install.packages("neuralnet")
set.seed(2)
dataset.tR  <- Smarket.tR
response.tR <- Direc.tR
dataset.tS  <- Smarket.tS
response.tS <- Direc.tS


  NN.fit = neuralnet(Direction ~ Lag1 + Lag2,
                     dataset.tR, hidden = 20, linear.output=FALSE)
          # lin.out=FALSE means activation function is applied to output node
# plot(NN.fit)


    # Predict for TEST set
    NN.pred = compute(NN.fit,  dataset.tS[, 1:8])


library(caret)
CM <- confusionMatrix(factor(NN.pred), factor(Direction.tS), positive="Up")
CM


    plot.roc(response.tS, as.vector(NN.pred$net.result[,1]), levels=c("Down", "Up"))
    abline(h=.7798, v=.9348, col="red")   #  point out specific place

    auc(response.tS, as.vector(NN.pred$net.result[,1]), levels=c("Down", "Up"))








###-------- Fit Neural Network
library(neuralnet)          # install.packages("neuralnet")
set.seed(2)
dataset.tR  <- Smarket.tR
response.tR <- Direc.tR
dataset.tS  <- Smarket.tS
response.tS <- Direc.tS


  NN.fit = neuralnet(Direction ~ Lag1 + Lag2 + Volume,
                     dataset.tR, hidden = 20, linear.output=FALSE)
          # lin.out=FALSE means activation function is applied to output node
# plot(NN.fit)


    # Predict for TRAINING set
    NN.pred.all = compute(NN.fit,  dataset.tR[, 2:8])
    NN.pred     = NN.pred.all$net.result[,1]
    NN.pred.UD = ifelse(NN.pred>.5, "Down", "Up")

    library(caret)
    CM <- confusionMatrix(factor(NN.pred.UD), factor(response.tR), positive="Up")
    CM

    library(pROC)
    plot.roc(response.tR, as.vector(NN.pred), levels=c("Down", "Up"))

    auc(response.tR, as.vector(NN.pred), levels=c("Down", "Up"))



    # Predict for TEST set
    NN.pred.all = compute(NN.fit,  dataset.tS[, 2:8])
    NN.pred     = NN.pred.all$net.result[,1]
    NN.pred.UD = ifelse(NN.pred>.5, "Down", "Up")

    library(caret)
    CM <- confusionMatrix(factor(NN.pred.UD), factor(response.tS), positive="Up")
    CM

    library(pROC)
    plot.roc(response.tS, as.vector(NN.pred), levels=c("Down", "Up"))
    abline(h=.7798, v=.9348, col="red")   #  point out specific place

    auc(response.tS, as.vector(NN.pred), levels=c("Down", "Up"))