Hello,
Is it possible to build a strategy using machine learning? How can I do that?
Thank you!
Hi,
Below an example of a strategy on the CAC relying on the popular XGBoost.
The strategy is a classification problem. The objective is to forecast the 3-step ahead returns going-up or down.
When the model forecasts a bullish trend, the strategy generates a buy signal (1). At the opposite, the signal is 0 and no position is taken.
To build my features, I used well-known technical indicators: Exponential Moving Average, Relative Strength Index and returns at different time horizons.
The EMA is a smoothing indicator showing price trend direction. The RSI on the other hand, is an oscillator measuring the speed and change of price movements.
I have trained my model on period ranging from 2007 to 2014 and tested it strating from 2015. I have obtained my performance results relying on Alphien’s algoEngine. Note that my model is trained within my payout function XGBoostStrat.
Feel free to copy this payout and modify it to suit your needs.
Thank you!
.sourceQlib()
XGBoostStrat<-function(ind=“CACa”, zoom=“2014”, forecastHorizon=3){
##Retrieving data
data=getLyxorBBs(ind,asPrice = TRUE)
##Building Features
features=cbind(data,
EMA(x=data,n=5),EMA(x=data,n=10), ##Exponential Moving Average
RSI(price=data, n=5), RSI(price=data, n=60), ##Relative Strength Index
ROC(data,type=“discrete”,n=2,na.pad=FALSE),
ROC(data,type=“discrete”,n=5,na.pad=FALSE))
features=na.omit(cbind(features, lag(features,2), lag(features,5), lag(features,10)))
##Set-up of the classification problem: Forcasting 5-step ahead returns going up (1) or down (0)
Y=ifelse(lag(ROC(data,type=“discrete”,n=forecastHorizon,na.pad=FALSE),-forecastHorizon)>=0, 1, 0)
data=na.omit(cbind(features,Y))
##Split data into train/test sets
data=list(“train”=data[paste0("/",zoom)], ‘test’=data[paste0((zoom+1),"/"),])
##Scaling Model
trainMeans=apply(data$train[,1:(ncol(data$train)-1)],2,mean) ##Estimating mean
trainSd=apply(data$train[,1:(ncol(data$train)-1)],2,sd) ##Estimating Standard deviation
train= as.data.frame(t(apply(data$train[,1:(ncol(data$train)-1)],1,function(x){(x-trainMeans)/trainSd})))##Normalize Train set
test = as.data.frame(t(apply(data$test[,1:(ncol(data$test)-1)],1,function(x){(x-trainMeans)/trainSd})))##Normalize Test set with train set metrics
##Training Model
model=xgboost::xgb.train(params=list(“max.depth” = 20, “eta” =0.1,
“min_child_weight” = 1, objective = “binary:logistic”),
data=xgb.DMatrix(data = as.matrix(train),
label = as.vector(data$train[,(ncol(train)+1)])),
nrounds=10000)
##Returning Payout: signals are the predictions
return(xts(ifelse(predict(model, (xgb.DMatrix(data = as.matrix(test))))>=0.5,1,0), order.by=index((data$test))))
}
algoEngine(“CACa”,type = “MM”) %>%
payout(XGBoostStrat, ind=“CACa”, zoom=2014) -> myXGBoostStrat
myXGBoostStrat=evaluate(myXGBoostStrat)
analyse(myXGBoostStrat)
backtest(myXGBoostStrat)
Post deleted by manas