Genetic Algorithm in R – Trend Following

This post is going to explain what genetic algorithms are, it will also present R code for performing genetic optimisation.

A genetic algo consists of three things:

  1. A gene
  2. A fitness function
  3. Methods to breed/mate genes

The Gene

The gene is typically a binary number, each bit in the binary number controls various parts of your trading strategy. The gene below contains 4 sub gene, a stock gene to select what stock to trade, a strategy gene to select what strategy to use, paramA sets a parameter used in your strategy and paramB sets another parameter to use in your strategy.

Gene = [StockGene,StrategyGene,ParamA,ParamB]

Stock Gene
00 Google
01 Facebook
10 IBM
11 LinkedIn

Strategy Gene
0 Simple Moving Average
1 Exponential Moving Average

ParamA Gene – Moving Average 1 Lookback
00 10
01 20
10 30
11 40

ParamB Gene – Moving Average 2 Lookback
00 15
01 25
10 35
11 45

So Gene = [01,1,00,11]

Would be stock=Facebook, strategy=Exponential Moving Average,paramA=10,paramB=45].

The strategy rules are simple, if the moving average(length=paramA) > moving average(length=paramB) then go long, and vice versa.

The fitness function

A gene is quantified as a good or bad gene using a fitness function. The success of a genetic trading strategy depends heavily upon your choice of fitness function and whether it makes sense with the strategies you intend to use. You will trade each of the strategies outlined by your active genes and then rank them by their fitness. A good starting point would be to use the sharp ratio as the fitness function.

You need to be careful that you apply the fitness function to statistically significant data. For example if you used a mean reverting strategy that might trade once a month (or what ever your retraining window is), then your fitness is determined by 1 or 2 datapoints!!! This will result in poor genetic optimisation (in my code i’ve commented out a mean reversion strategy test for yourself). Typically what happens is your sharpe ratio from 2 datapoints is very very high merely down to luck. You then mark this as a good gene and trade it the next month with terrible results.

Breeding Genes

With a genetic algo you need to breed genes, for the rest of this post i’ll assume you are breeding once a month. During breeding you take all of the genes in your gene pool and rank them according to the fitness function. You then select the top N genes and breed them (discard all the other genes they’re of no use).

Breeding consists of two parts:

Hybridisation – Take a gene and cut a chunk out of it, you can use whatever random number generator you want to determine the cut locations, swap this chunk with a corresponding chunk from another gene.

Old gene: 00110010 and 11100110 (red is the randomly select bits to cut)
New gene: 00100110 and 11110010

You do this for every possible pair of genes in your top N list.

Mutation – After hybridisation go through all your genes and randomly flip the bits with an fixed probability. The mutation prevents your strategy from getting locked into an every shrinking gene pool.

For a more detailed explanation with diagrams please see: scroll down to Genetic Algorithms and its Application in Trading

genetic algo sharpe 1.14

Annualized Sharpe Ratio (Rf=0%) 1.15

On to the code:

?View Code RSPLUS
topNToSelect <- 5   #Top n genes are selected during the mating, these will be mated with each other
mutationProb <- 0.05 #A mutation can occur during the mating, this is the probability of a mutation for individual chromes
symbolLst <- c("^GDAXI","^FTSE","^GSPC","^NDX","AAPL","ARMH","JPM","GS")
#Stock gene
stockGeneLength <- 3 #8stocks
#stockGeneLength<-6 #Allows 2^6 stocks (64)
#Strategy gene
#Paramter lookback gene
#Calculate the length of our chromozone, chromozone=[gene1,gene2,gene3...]
chromozoneLength <- stockGeneLength+strateyGeneLength+parameterLookbackGeneLength
signalMACross <- function(mktdata, paramA, paramB, avgFunc=SMA){
 signal = avgFunc(mktdata,n=paramA)/avgFunc(mktdata,n=paramB)
 signal[] <- 0
 signal <- (signal>1)*1 #converts bools into ints
 signal[signal==0] <- (-1)
 return (signal)
signalBollingerReversion <- function(mktdata, paramA, paramB){
  avg <- SMA(mktdata,paramB)
  std <- 1*rollapply(mktdata, paramB,sd,align="right")
  shortSignal <- (mktdata > avg+std)*-1
  longSignal <- (mktdata < avg-std)*1
  signal <- shortSignal+longSignal
  return (signal)
signalRSIOverBoughtOrSold <- function(mktdata, paramA, paramB){
  upperLim <- min(60*(1+paramB/100),90)
  lowerLim <- max(40*(1-paramB/100),10)
  rsisignal <- RSI(mktdata,paramB)
  signal <- ((rsisignal>upperLim)*-1)+((rsisignal<lowerLim)*1)
  return (signal)
#Gene = [StockGene,StrategyGene,ParamAGene,ParamBGene]
#The following functions extract specific parts of the gene
getStockGeneFromChromozone <- function(chrome){
getStrategyGeneFromChromozone <- function(chrome){
getParameterLookbackGeneFromChromozone <- function(chrome){
#Once parts of the gene have been extracted they are then converted into
#lookback values, what stocks to trade, or what strategy to use
getStockDataFromChromozone<- function(chrome){
      #Basically a binary number to decimal converter
      gene <- getStockGeneFromChromozone(chrome)
      index <-sum(chrome*(2^(seq(1,length(gene),1)-1)))+1 #The +1 is to stop 0 since not a valid index
      cleanName <- sub("^","",symbolLst[index],fixed=TRUE)
     return (get(cleanName,symbolData))
getStrategyFromChromozone <- function(chrome){
      gene <- matrix(getStrategyGeneFromChromozone(chrome))
       return (signalMACross)
       return (function(mktdata,paramA,paramB) {signalMACross(mktdata,paramA,paramB,avgFunc=EMA)})
       return (function(mktdata,paramA,paramB) {signalMACross(mktdata,paramA,paramB,avgFunc=ZLEMA)})
      # return (signalBollingerReversion)
       return (function(mktdata,paramA,paramB) {signalMACross(mktdata,paramA,paramB,avgFunc=WMA)})
      # return (signalRSIOverBoughtOrSold)
      print("nothing found")
getLookbackAFromChromozone <- function(chrome){
    gene <- getParameterLookbackGeneFromChromozone(chrome)
    gene <- gene[,1:3]
    gene <- matrix(gene)
        return (10)
        return (20)
        return (30)
        return (40)
        return (50)
        return (60)
        return (70)
        return (80)
getLookbackBFromChromozone <- function(chrome){
    gene <- getParameterLookbackGeneFromChromozone(chrome)
    gene <- gene[,4:6]
    gene <- matrix(gene)
        return (15)
        return (25)
        return (35)
        return (45)
        return (55)
        return (65)
        return (75)
        return (85)
#The more positive the fitness, the better the gene
calculateGeneFitnessFromTradingReturns <- function(tradingRet){
  tradingFitness <- SharpeRatio.annualized(tradingRet)
  #tradingFitness <- SharpeRatio.annualized(tradingRet) * (1/maxDrawdown(tradingRet))
  #tradingFitness <- max(cumsum(tradingRet))/maxDrawdown(tradingRet)
  #tradingFitness <- sum((tradingRet>0)*1)/length(tradingRet) #% of trades profitable
  #tradingFitness <- -1*maxDrawdown(tradingRet)
#This function performs the mating between two chromozones
genetricMating <- function(chromozoneFitness,useTopNPerformers,mutationProb){
        selectTopNPerformers <- function(chromozoneFitness,useTopNPerformers){
              #Ranks the chromozones by their fitness and select the topNPerformers
              orderedChromozones <- order(chromozoneFitness[,"Fitness"],decreasing=TRUE)
              orderedChromozones <- chromozoneFitness[orderedChromozones,]
              ##Often there are lots of overlapping strategies with the same fitness
              ##We should filter by unique fitness to stop the overweighting of lucky high fitness
              orderedChromozones <- subset(orderedChromozones, !duplicated(Fitness))
        hybridize <- function(topChromozones,mutationProb){
            crossoverFunc <- function(chromeA,chromeB){
            chromeA <- chromeA[,!colnames(chromeA) %in% c("Fitness")]
            chromeB <- chromeB[,!colnames(chromeB) %in% c("Fitness")]
                  #Takes a number of chromes from B and swaps them in to A
                  nCross <- runif(min=0,max=ncol(chromeA)-1,1) #the number of individual chromes to swap
                  swapStartLocation = round(runif(min=1,max=ncol(chromeA),1))
                  swapLocations <- seq(swapStartLocation,swapStartLocation+nCross) #Can run over the end of our vector, need to wrap around back to start
                  swapLocations <- swapLocations %% ncol(chromeA)+1 #Performs the wrapping
                  chromeA[1,swapLocations] <- chromeB[1,swapLocations] #Performs the swap
                  return (chromeA)
            mutateFunc <- function(chrome,mutationProb){
                return((round(runif(min=0,max=1,ncol(chrome))<mutationProb)+chrome) %% 2)
            #Take each chromozone and mate it with all the others (and it's self)
            a <- topChromozones[rep(seq(1,nrow(topChromozones)),each=nrow(topChromozones)),] #Repeat each row nrow times
            b <- topChromozones[rep(seq(1,nrow(topChromozones)),nrow(topChromozones)),] #Repeat whole matrix nrow times
            #Can this be vectorised (not huge amounts of data anyway so probs not an issue)?
            res <- matrix(nrow=0,ncol=ncol(a)-1) #The minus 1 is to drop the "Fitness" column
            for(i in 1:nrow(a)){
                res <- rbind(res,mutateFunc(crossoverFunc(a[i,],b[i,]),mutationProb))
            return (res)
        topChromozones <- selectTopNPerformers(chromozoneFitness,useTopNPerformers)
        #return ((hybridize(topChromozones,mutationProb))) #You may want duplicates to give more weight to 'good' genes
        return (unique(hybridize(topChromozones,mutationProb))) #Remove duplicate genes
#This function takes a chrome/gene and does the according trades
#It takes market data and a start and an end date
#It does not take responsibility for the mating and ranking of genes
doGeneticTrading <- function(mktdata,chrome, startDate, endDate){
    signalFunc <-getStrategyFromChromozone(chrome)
    paramA <- getLookbackAFromChromozone(chrome)
    paramB <- getLookbackBFromChromozone(chrome)
    signal <- signalFunc(Op(mktdata),paramA,paramB)
    opClRet <- (Cl(mktdata)/Op(mktdata)) - 1
    tradingReturns = opClRet * signal
    dataWin <- (paste(startDate,"::",endDate,sep=""))
    tradingReturns <- tradingReturns[dataWin]
    colnames(tradingReturns) <- c("TradingRet")
#This function mates genes every month
#It also passes those genes into the doGeneticTrading function
doTrading <- function(chromelist){
  #Function for taking a year and a month and spitting out a clean date
  cleanDate <- function(y,m){
      if(m == 13){
       m <- 1
       y <- y+1
       if(m < 10){
       } else {
  year <- 2002
  month <- 1
  totalRet <- 0
  fitnessEvoltion <- 0
  #Loop through many years and months
  for(y in 2002:2010){
    for(m in 1:12){
        chromeFitness <-,ncol=ncol(chromelist)))
        startD <- cleanDate(y-2,m) #Subtracting off 2 years to ensure we pass enough data in(should really be calculated from MA lookback)
        liveStart <- cleanDate(y,m)
        liveEnd <- cleanDate(y,m+1)
        dataWin <- (paste(startD,"::",liveEnd,sep=""))
        monthReturn <- data.frame()
        #Look through all the active chromes and use them for trading
        for(cn in 1:nrow(chromelist)){
        #USE a try catch just incase there are data issue etc...
            mktdata <- getStockDataFromChromozone(chromelist[cn,])
            tradingRet <- doGeneticTrading(mktdata[dataWin],chromelist[cn,],liveStart,liveEnd)
            tradingRet <- tradingRet*(1/nrow(chromelist)) #even money given to each strategy
            tradingFitness <- calculateGeneFitnessFromTradingReturns(tradingRet)
            if(!is.nan(tradingFitness) && !is.nan(max(tradingRet)) && !is.nan(min(tradingRet))){
              if(length(monthReturn) == 0 ){
               monthReturn <- tradingRet
              } else {
               monthReturn <- cbind(monthReturn,tradingRet)
            res <- cbind(chromelist[cn,],tradingFitness)
            colnames(res) <- c(colnames(chromelist[cn,]),"Fitness")
            chromeFitness <- rbind(chromeFitness,res)
        print("Month return")
        #Collapse all the trades from each chromozone into a single P&L for each day in the month
        monthReturn <- apply(monthReturn,1,sum,na.rm=TRUE)
        currentMonthFitness <- calculateGeneFitnessFromTradingReturns(monthReturn)
        #Update the running total of P&L
        totalRet <- c(totalRet,monthReturn)
        fitnessEvoltion <- c(fitnessEvoltion,currentMonthFitness)
        chromelist <- genetricMating(chromeFitness,topNToSelect,mutationProb)
        print(paste("There are",nrow(chromelist), "chromes active"))
        print(paste("Min Fitness:",min(chromeFitness[,"Fitness"])))
        print(paste("Max Fitness:",max(chromeFitness[,"Fitness"])))
        print(paste("Average Fitness:",mean(chromeFitness[,"Fitness"])))
        print(paste("Current Month Fitness:",currentMonthFitness))
   return (totalRet)
#Specify dates for downloading data, training models and running simulation
startDate = as.Date("2000-01-01") #Specify what date to get the prices from
symbolData <- new.env() #Make a new environment for quantmod to store data in
getSymbols(symbolLst, env = symbolData, src = "yahoo", from = startDate)
#Create some genes at random
#Make a diag matrix so that each chrome gets activated atleast once
startingChromozones <- diag(chromozoneLength)
rownames(startingChromozones) <- apply(t(seq(1,chromozoneLength)),2,function(x) { paste("Chrome",x,sep="") } )
fitness <- matrix(runif(min=-1,max=1,nrow(startingChromozones)),nrow=nrow(startingChromozones),ncol=1)
colnames(fitness) <- c("Fitness")
startingChromozones <-,fitness))
print("Before mating")
print("After mating")
startingChromozones <- genetricMating(startingChromozones,topNToSelect,mutationProb)
tradingReturns <- doTrading(startingChromozones)
tradingReturns <-[-1])))
charts.PerformanceSummary(tradingReturns,main=paste("Arithmetic Genetic Trading Returns"),geometric=FALSE)
cat("Sharpe Ratio")