This is an example code for BOUN58D project.

example_model_with_CART.R

#Get the required libraries, install first if not installed.
options(dplyr.width=Inf)
library(boun58d) #Copy paste the necessary function if you cannot install this from github
library(tidyverse)
library(tidyquant)
library(rpart)
library(rpart.plot)
library(rattle)
set.seed(58) #Don't forget to set the seed for reproducibility

#Add some commments to your code so others would understand

start_time = as.numeric(as.POSIXct("2015-08-01  00:00:00",tz="UTC"))
end_time = as.numeric(as.POSIXct("2017-04-01  00:00:00",tz="UTC"))
period_interval <- 86400 #Daily
lookback_period <- 365
the_currency <- "BTC_ETH"
#Get the Data
chartData <- get_from_poloniex("returnChartData",parameters=list(currencyPair = the_currency,start=start_time,end=end_time,period=period_interval))
#Convert chartData into a data.table data frame with data.table::rbindlist
chartData <- data.table::rbindlist(chartData)

analysis_df <-
chartData %>%
transmute(numdate = date,date=as.POSIXct(date,origin="1970-01-01",tz="UTC"),close=close) %>%
mutate(logreturn=log(close/lag(close,1)) #calculate log-returns
) %>%
filter(!is.na(logreturn)) %>% #remove the NA generated by first calculation
mutate(
    ma5=rollmean(logreturn,5,align="right",fill=NA), #calculate MA5
    ma20=rollmean(logreturn,20,align="right",fill=NA) #calculate MA20
) %>%
tbl_df

#Arrange the data frame to suit with forecasting needs
#Move predictors to the to be predicted values so we can add the predictions right next to actual values
forecast_df <-
analysis_df %>%
mutate(s0=lag(close,1),ma5=lag(ma5,1),ma20=lag(ma20,1)) %>% #move predictors one day ahead
filter(complete.cases(.)) %>% #remove NA
arrange(desc(numdate)) #Rearrange by descending date

#Determine the dates to forecast and select predictors
oos_forecast_df <- forecast_df %>% slice(1:30) %>% select(date,ma5,ma20)

#Initiate the predicted log-return variable
pred_logreturn <- 0

for(i in 1:30){
    #Run the model with the available information for each day
    rpart_model <- rpart(logreturn ~ ma5 + ma20,data= forecast_df %>% filter(date < oos_forecast_df$date[i]))
    #Get the logreturn prediction as an output of the model
    pred_logreturn[i] <- predict(rpart_model,newdata=oos_forecast_df %>% slice(i) %>% select(ma5,ma20))
}

#Merge the findings with our forecast_df
results_df<-
oos_forecast_df %>%
transmute(date,pred_logreturn=pred_logreturn) %>%
left_join(forecast_df,.,by="date") %>%
mutate(
    pred_close = s0*exp(pred_logreturn),
    position=ifelse(pred_close>=s0,"long","short"),
    rpe=round((pred_close-close)/close,4),
    ape=abs(pred_close-close),
    pnl=exp(logreturn)^ifelse(pred_close >= s0,1,-1)
) %>%
tbl_df

#This graph is prediction vs realization
ggplot(data=results_df %>% slice(1:50), aes(x=date)) +
geom_line(aes(y=close)) +
geom_line(aes(y=pred_close),color="red") +
labs(title="Prediction vs Realization")

#This graph is ape
ggplot(data = results_df %>% filter(!is.na(ape)),aes(x=date,y=ape)) + geom_line() +
labs(title="Absolute Pricing Error")
#This graph is rpe
ggplot(data = results_df %>% filter(!is.na(rpe)),aes(x=date,y=rpe)) + geom_line() +
labs(title="Relative Pricing Error")
#This graph is pnl
ggplot(data = results_df %>% filter(!is.na(pnl)),aes(x=date,y=pnl)) + geom_line() +
labs(title="Daily Profit and Loss")

cumulative_pnl_data <-
results_df %>%
    filter(!is.na(pnl)) %>%
    transmute(date,pnl) %>%
    add_row(date=min(.$date - 86400),pnl=1) %>%
    arrange(date) %>%
    mutate(cum_pnl = cumprod(pnl)*100)

ggplot(data=cumulative_pnl_data,aes(x=date,y=cum_pnl)) + geom_line() +
labs(title="Cumulative PnL if started with 100 BTC/ETH")