Neural network test in R

# using Artificial Neural Networks to predict (unsuccessfully...) stock price changes
# see more: http://reakkt.com

rm(list=ls())

library(neuralnet)

data.file <- "fw20_d.csv" # EOD data from stooq.com

# n=500, d=50, l=4 lub 3 dla t=10
# n=500, d=50, l=3 dla t=5

d <- 50

n.train <- 500
n.test  <- 7

layers      <- 3
repetitions <- 2

n <- n.train+n.test+d

fit.n <- 100 # number of s1 date draws

market.data <- read.csv(data.file,header=TRUE)

x.market <- log(market.data[-1,"Close"]/
                market.data[-nrow(market.data),"Close"])

market.n <- length(x.market)

fit.ratio  <- numeric(fit.n)
random.fit <- numeric(fit.n)

for (i in 1:fit.n) {

  s1 <- round(runif(1,n,market.n))

  x <- x.market[(s1-n+1):s1]

  x.matrix <- cbind( tail(x,n-d),
                     sapply(1:(d-1), function(i) tail(head(x,n-i),n-d) ),
                     head(x,n-d) )

  colnames(x.matrix) <- c("X",paste("X",1:d,sep=""))
  x.df <- data.frame(x.matrix)

  x.df.train <- x.df[1:n.train,]
  x.df.test  <- x.df[(n.train+1):(n.train+n.test),]

  # some other parameters to try:
  # act.fct="tanh"
  # err.fct="ce"

  f <- paste("X",paste("X",1:d,"+",sep="",collapse=""),sep="~")
  f <- substr(f,1,nchar(f)-1)

  nn <- neuralnet(eval(parse(t=f)),data=x.df.train,
                  # act.fct="tanh",
                  hidden=layers,rep=repetitions,linear.output=TRUE)

  if (length(nn$weights)==0) stop("No convergence") else {

    # plot(nn)

    x.forecast <- compute(nn,x.df.test[,-1])

    matplot(cbind(x.df[(n-d-n.test+1):(n-d),1],x.forecast$net.result),type="l",lty="solid",ylab="")
    points(x.df[(n-d-n.test+1):(n-d),1],col="Black",pch=16)
    points(x.forecast$net.result,col="Red",pch=16)
    abline(h=0,col="Orange",lty="dotted")

    # cbind(x.df.test[,1],x.forecast$net.result)

    conform.sign <- sum(sign(x.df.test[,1])==sign(x.forecast$net.result))

    # conform.diff <- sum((x.df.test[,1]-x.forecast$net.result)^2)

  }

  fit.ratio[i]  <- conform.sign/n.test

  random.fit[i] <- sum(sign(x.df.test[,1])==sapply(runif(n.test,0,1), function(i) if (i>0.5) return(1) else return(-1)))/n.test

}

fit.density     <- density(fit.ratio)
random.density  <- density(random.fit)

xlim <- range(c(fit.density$x,random.density$x))
ylim <- range(c(fit.density$y,random.density$y))

plot(fit.density,xlim=xlim,ylim=ylim)
lines(random.density,col="Blue",lty="dotted")
abline(v=0.5,col="Red",lty="dotted")

summary(random.fit)
summary(fit.ratio)