Implemented the bandit strategy

not really working so far

1 files changed, 36 insertions, 78 deletions

diff --git a/bandit.R b/bandit.R
index 8e2fe70..4ce86bf 100644
--- a/bandit.R
+++ b/bandit.R
@@ -1,7 +1,9 @@
+source('utils.R')
 sp500<- read.table("sp500 tickers 10-17.csv",sep="\t",header=T,quote="",colClasses="character",strip.white=T)
 colnames(sp500) <- c("name","ticker")
 #replace / by -
 sp500$ticker <- sub('/', '-', sp500$ticker)
+load("sp500-bloomberg.RData")
 
 #first don't take care about survivor bias
 tickerlist <- list()
@@ -11,103 +13,59 @@ for(i in 1:length(sp500$ticker)){
   cat("ticker",i,ticker,"\n")
 }
 
-#compute matrix of returns
-ticker <- sp500[1,"ticker"]
-returns <- diff(log(tickerlist[[ticker]]$Close))
-for(i in 2:length(sp500[,"ticker"])){
-  ticker <- sp500[i,"ticker"]
-  returns <- merge(returns,diff(log(tickerlist[[ticker]]$Close)))
+price2return <- function(x){
+  diff(log(abs(x)))
+}
+
+#compute list of returns
+tickerlist <- names(stock.data)
+returns <- price2return(stock.data[[tickerlist[1]]]$Close)
+for(i in 2:length(tickerlist)){
+  ticker <- tickerlist[i]
+  returns <- merge(returns,price2return(stock.data[[ticker]]$Close),all=T)
 }
-colnames(returns) <- sp500[,"ticker"]
+colnames(returns) <- tickerlist
 returns <- returns[-1,]
-universe <- which(!is.na(returns[1,]))
-returns.subset <- returnds[,universe]
 
 #constant weight rebalancing
-w <- rep(1/dim(returns.subset)[2],dim(returns.subset)[2])
 N <- 1000000
 W <- N
-for(day in 1:length(returns[,1])){
-  r <- as.vector(exp(returns.subset[day,])-1)
+for(day in 1:NROW(returns)){
+  date <- time(returns)[day]
+  subset <- memb(sp500$ticker,add,date)
+  r <- as.vector(exp(returns[day,subset])-1)
   r[is.na(r)] <- 0 #NA means we don't have quotes that day, but next return
                    #will catch up
+  w <- rep(1/length(r),length(r))
   dN <- N*crossprod(w,r)
   N <- N+dN
   W <- c(W,N)
 }
-#no rebalancing
-N <- 1000000
-W <- N
-R <- rep(1,dim(returns.subset)[2])
-for(day in 1:length(returns[,1])){
-  r <- as.vector(exp(returns.subset[day,]))
-  r[is.na(r)] <- 1
-  R <- R*r
-  W <- c(W,N*mean(R))
-}
-
-indexchange <- read.table("IndexChange_500.csv",sep="\t",header=T,quote="",
-                          colClasses="character")
-#cleanup tickers
-indexchange$Ticker.add <- sub('/', '-',indexchange$Ticker.add)
-indexchange$Ticker.add <- sub('.wi', '',indexchange$Ticker.add,fixed=T)
-indexchange$Ticker.add <- sub('wi', '',indexchange$Ticker.add)
-indexchange$Ticker.add <- sub('.', '-',indexchange$Ticker.add,fixed=T)
-indexchange$Ticker.del <- sub('/', '-',indexchange$Ticker.del)
-indexchange$Ticker.del <- sub('.wi', '',indexchange$Ticker.del,fixed=T)
-indexchange$Ticker.del <- sub('wi', '',indexchange$Ticker.del)
-indexchange$Ticker.del <- sub('.', '-',indexchange$Ticker.del,fixed=T)
-
-#rollback sp500 changes
-sp500.historic <- sp500
-for( i in 1:100){
-  #Ticker.add empty means index is <500 for a while
-  if(indexchange$Ticker.add[i]!=""){
-    if(!is.na(match(indexchange$Ticker.add[i],sp500.historic$ticker))){
-      sp500.historic <-
-        sp500.historic[-match(indexchange$Ticker.add[i],sp500.historic$ticker),]
-    }else{
-      cat(paste("ticker",indexchange$Ticker.add[i],"not found\n"))
-      next()
-    }
-  }
-   #add ticker that was removed
-  if(indexchange$Ticker.del[i]!=""){
-    sp500.historic <- rbind(sp500.historic,
-                            c(indexchange$Company.Deletions[i],
-                              indexchange$Ticker.del[i]))
-  }
-}
 
-n.stocks <- ncol(P)
-N <- nrow(P)
+#bandit strategy
 current.wealth <- 1
-w <- rep(1/n.stocks,n.stocks)
-dP <- apply(P,2,diff)
-L <- rep(0,n.stocks)
 V <- 0
-W <- w
 pnl <- 0
-for(i in 1:(N-1)){
-  r <- dP[i,]/as.numeric(P[i,])
-  #r <- c(r,-r)
-  pnl <- cbind(pnl,current.wealth*crossprod(w,r))
-  current.wealth <- 1+sum(pnl)
-  L <- rbind(L,-r)
-  V <- V+crossprod(w,r^2)
-  T <- 2/3*sqrt(log(N)/V)
-  #w <- exp(-T*(apply(1+L,2,prod)-1))
-  w <- exp(-T*colSums(L))
+members <- memb(sp500$ticker,add,as.Date("2000-01-01"))
+w <- rep(1/499,499)
+L <- rep(0,NCOL(returns))
+for(day in 1:NROW(returns)){
+  #browser()
+  date <- time(returns)[day]
+  r <- exp(coredata(returns[day,]))-1
+  r[is.na(r)] <- 0
+  V <- V+drop(crossprod(w,r[,members]^2))
+  #L <- exp(-colSums(returns[1:day,],na.rm=T))-1
+  L <- L-r
+  pnl <- c(pnl, current.wealth*crossprod(w,r[,members]))
+  current.wealth <- 1 + sum(pnl)
+  members <- memb(sp500$ticker,add,date)
+  #eta <- 2/3*sqrt(log(500)/V)
+  eta <- 0.15
+  w <- exp(-eta*L[,members])
   w <- w/sum(w)
-  W <- rbind(W,w)
-  if(i%%10==0){
-    cat(current.wealth,sep="\n")
-  }
 }
 
-price2return <- function(x){
-  diff(x)/x[-length(x)]
-}
 #number of shares implementations
 tc <- 0.005+0.02 
 days <- nrow(data.bus)