Goomboo's EMA10/21 Crossover Versus the rebuilder's Flipist Method

[stochastic]

Now for the Flipist strategy.  This is the same strategy as before but now we can flip every hour if needed.

First the code.  I added in everything to this code, from making the signals, making the charts, and outputting the statistics.

Code:

############################################
#Backtesting Flipist Method
############################################
    # We will need the quantmod package for charting and pulling
    # data and the TTR package to calculate RSI(2).
    # You can install packages via: install.packages("packageName")
    # install.packages(c("quantmod","TTR"))
    library(quantmod)
    library(TTR)

    # Load data
    x = last(y,4415) #gets the last 184 daily positions from the data (starts at July 23, 2011)

    # Calculate the random indicator
    set.seed(43) #include this to reproduce my results
    x$flip <- rbinom(4415,1,0.5) #50% probability that 1 (heads/buy) will show and 50% probability that 0 (tails/sell) will show

    # Create the buy (1) and sell (-1) signals
    sigbuy <- ifelse(x$flip == 1, 1, 0)
    sigsell <- ifelse(x$flip == 0, -1, 0)

    # Lag signals to align with days in market,
    # not days signals were generated
    sigbuy <- lag(sigbuy,1) # Note k=1 implies a move *forward*
    sigsell <- lag(sigsell,1) # Note k=1 implies a move *forward*

    # Replace missing signals with no position
    # (generally just at beginning of series)
    sigbuy[is.na(sigbuy)] <- 0
    sigsell[is.na(sigsell)] <- 0

    # Combine both signals into one vector
    sig <- sigbuy + sigsell
    # Calculate Close-to-Close returns
    ret <- ROC(Cl(x))
    ret[1] <- 0

    # Calculate equity curves
    eq_up <- exp(cumsum(ret*sigbuy))
    eq_dn <- exp(cumsum(ret*sigsell*-1))
    eq_all <- exp(cumsum(ret*sig))

    # Equity Chart
    png("flipist.png",width=720,height=720)
    plot.zoo( cbind(eq_up, eq_dn),
    ylab=c("Long","Short"), col=c("green","red"),
    main="Flipist Strategy: 07-23-2011 to 01-22-2012" )
    dev.off()

    # Create a chart showing mtgoxUSD
    png("flipistchart.png",width=720,height=720)
    chartSeries(x, subset="last 4415 hours", type="line")

    # Add the total equity line
    addTA(eq_all)
    dev.off()

    # Evaluate the Strategy

    # install.packages("PerformanceAnalytics")
    require(PerformanceAnalytics)
    
    # chart equity curve, daily performance, and drawdowns
    png("performance.png",height=720,width=720)
    charts.PerformanceSummary(ret)
    dev.off()

    # This function gives us some standard summary
    # statistics for our trades.
    tradeStats <- function(signals, returns) {
    # Inputs:
    # signals : trading signals
    # returns : returns corresponding to signals

    # Combine data and convert to data.frame
    sysRet <- signals * returns * 100
    posRet <- sysRet > 0 # Positive rule returns
    negRet <- sysRet < 0 # Negative rule returns
    dat <- cbind(signals,posRet*100,sysRet[posRet],sysRet[negRet],1)
    dat <- as.data.frame(dat)

    # Aggreate data for summary statistics
    means <- aggregate(dat[,2:4], by=list(dat[,1]), mean, na.rm=TRUE)
    medians <- aggregate(dat[,3:4], by=list(dat[,1]), median, na.rm=TRUE)
    sums <- aggregate(dat[,5], by=list(dat[,1]), sum)

    colnames(means) <- c("Signal","% Win","Mean Win","Mean Loss")
    colnames(medians) <- c("Signal","Median Win","Median Loss")
    colnames(sums) <- c("Signal","# Trades")

    all <- merge(sums,means)
    all <- merge(all,medians)

    wl <- cbind( abs(all[,"Mean Win"]/all[,"Mean Loss"]),
    abs(all[,"Median Win"]/all[,"Median Loss"]) )
    colnames(wl) <- c("Mean W/L","Median W/L")

    all <- cbind(all,wl)
    return(all)
    }

    print(tradeStats(sig,ret))

[stochastic]

As expected with a random method with a greater sample size (number of trades) it is almost even between buy and sell signal.

Code:

  
-1 = short
1 = long
Signal # Trades    % Win Mean Win Mean Loss Median Win Median Loss Mean W/L Median W/L
1     -1     2208 50.00000 1.123302 -1.034769  0.6283153  -0.6485819 1.085558  0.9687524
2      0        1  0.00000      NaN       NaN         NA          NA      NaN         NA
3      1     2206 49.50136 1.063690 -1.027253  0.6050835  -0.5874410 1.035470  1.0300328

[stochastic]

You need emotional investor and EMA10/21 contrarian controls. I'm not sure how to implement the emotional investor one...

How would you do the controls then?

[stochastic]

***EDIT*** Calculated Expectancy

Expectancy = (Probability of Win * Average Win) – (Probability of Loss * Average Loss)

Flipist Daily Flip

Code:

  (1) buy / (-1) sell
   Signal # Trades  % Win   Mean Win Mean Loss Median Win Median Loss Mean W/L Median W/L
1     -1       93   52.68817 7.132123 -6.507944   3.636423   -4.839239 1.095910  0.7514452
2      0        1    0.00000      NaN       NaN         NA          NA      NaN         NA
3      1       90   46.66667 4.551028 -4.431499   2.366660   -2.242254 1.026973  1.0554826

Flipist - Hourly Flip

Code:

 (1) buy / (-1) sell
    Signal # Trades  % Win   Mean Win Mean Loss Median Win Median Loss Mean W/L Median W/L
1     -1     2208   50.00000 1.123302 -1.034769  0.6283153  -0.6485819 1.085558  0.9687524
2      0        1    0.00000      NaN       NaN         NA          NA      NaN         NA
3      1     2206   49.50136 1.063690 -1.027253  0.6050835  -0.5874410 1.035470  1.0300328

Expectancy = -4.0892

EMA10/21 Crossover using Daily data.  There are 18 less trades when looking at the

Code:

  EMA10/21, buy (1) when EMA10 > EMA21, sell (-1) when EMA10 < EMA21
    Signal # Trades  % Win   Mean Win  Mean Loss  Median Win Median Loss Mean W/L Median W/L
1     -1      132    56.81818 6.341372 -5.713935 2.840616 -4.183966  1.109808 0.678929
2      0       21    0.00000      NaN       NaN       NA        NA       NaN       NA
3      1       52    57.69231  5.068756 -3.911978 3.199519 -1.747152  1.295702 1.831277

EMA10/21 Crossover using Hourly data.

Code:

 EMA10/21 Buy (1) when EMA10 crosses over EMA21, sell (-1) when EMA10 crosses below EMA21
    Signal # Trades    % Win   Mean Win Mean Loss Median Win Median Loss  Mean W/L Median W/L
1     -1       83    51.80723 0.8337976 -0.8799018  0.4516124  -0.5958119 0.9476030  0.7579782
2      0     4249     0.00000       NaN        NaN         NA          NA       NaN         NA
3      1       83    46.98795 0.6240971 -1.4239146  0.4776158  -0.5503347 0.4382968  0.8678642

Expectancy = -45.00281

[stochastic]

Using modified code from here and here.  I ran some backtesting in R to compare the EMA 5 EMA 21 Crossover and a random buy/sell strategy based on a coin flip.

Why did I do this?  I was making my own trading strategies but got stuck on making one for bullish/bearish divergence and I thought some practice may help me find out what I am doing wrong.

Why not generate your own? The R package rgp lets you evolve GAs. I've got some that consistently give me better than 'buy and hold'.

In artificial intelligence, genetic programming (GP) is an evolutionary algorithm-based methodology inspired by biological evolution to find computer programs that perform a user-defined task. It is a specialization of genetic algorithms (GA) where each individual is a computer program. It is a machine learning technique used to optimize a population of computer programs according to a fitness landscape determined by a program's ability to perform a given computational task.

Wow, thanks for the recommendation.  I am an evolutionary biologist (or trying to be) and this makes a lot of sense.  How have you applied GP to trading strategies?

[miscreanity]

Very nice work.

Also try expanding the time frame to daily & weekly scales.

[organofcorti]

Sure - although tbh I've only been using rgp for a couple of weeks, before that I was using eureqa but I didn't find it flexible enough (it's really just for symbolic regression).

rgp is well documented but if you don't have a lot of experience with GAs (i don't) then the tutorials at rsymbolic.org are a bit incomplete. Still, I have plenty of algorithms generated. Some overfit and just buy below a certain level and sell above it. It all comes down to your fitness function and i think mine leaves a bit to be desired.

Some of the useful trading algo's I've got are of the longlag-shortlag=0 variety, just like the EMA10 EMA21 variety.

I'll be happy to share code I've done so far - I'm sure you can improve on it.

[stochastic]

Sure - although tbh I've only been using rgp for a couple of weeks, before that I was using eureqa but I didn't find it flexible enough (it's really just for symbolic regression).

rgp is well documented but if you don't have a lot of experience with GAs (i don't) then the tutorials at rsymbolic.org are a bit incomplete. Still, I have plenty of algorithms generated. Some overfit and just buy below a certain level and sell above it. It all comes down to your fitness function and i think mine leaves a bit to be desired.

Some of the useful trading algo's I've got are of the longlag-shortlag=0 variety, just like the EMA10 EMA21 variety.

I'll be happy to share code I've done so far - I'm sure you can improve on it.

Thanks, if you can put some here or pm to me I will take a look at it.  I requested some books on the subject.

[stochastic]

Very nice work.

Also try expanding the time frame to daily & weekly scales.

I already have the daily scale up.  I personally think bitcoin is too volatile to really hold it long term in large quantities, so holding a position for large amounts for weeks or months would be unrealistic if you are using margin.

[miscreanity]

I already have the daily scale up.  I personally think bitcoin is too volatile to really hold it long term in large quantities, so holding a position for large amounts for weeks or months would be unrealistic if you are using margin.

Right now and for some time to come, yes. You'd have to manage a balanced position by scaling it instead of going all in each time.

With Bitcoin, I use weekly data as an overall trending and reversal guide, then daily/hourly to spot scaling opportunities. It'd be more complex using multiple temporal dimensions, but even building on the simple EMA10/21 technique that way would be far more robust - even somewhat predictive.

The greater the time-frame, the harder it is to reverse the trend.

[rebuilder]

Just popped in to say "nice work!"

I don't have the time to look through the whole thread right now, but I'll be sure to do it when I can.

[gamer4156]

It says EMA5/21 on the title of the chart?

[stochastic]

It says EMA5/21 on the title of the chart?

Thanks for catching that.  It should be EMA10

[stochastic]

I already have the daily scale up.  I personally think bitcoin is too volatile to really hold it long term in large quantities, so holding a position for large amounts for weeks or months would be unrealistic if you are using margin.

Right now and for some time to come, yes. You'd have to manage a balanced position by scaling it instead of going all in each time.

With Bitcoin, I use weekly data as an overall trending and reversal guide, then daily/hourly to spot scaling opportunities. It'd be more complex using multiple temporal dimensions, but even building on the simple EMA10/21 technique that way would be far more robust - even somewhat predictive.

The greater the time-frame, the harder it is to reverse the trend.

Next time I will add in scaling.  Anyone that wants to jump ahead check out here.  R has some other tools such as blotter that can keep track of positions.