Hail Traders!
Wondering if anyone can explain apparently wide fluctuations in test results over various baskets of instruments, over different time frames and with different start dates? For example, if a sytem consistently has 2 winners for every 3 losers and the winners are roughly 3 times the size of losers, then risk per trade becomes the key factor I believe. However, the results, while always positive, vary greatly, which I assume to be essentially "random". If that is the case, then isn't historical testing just "mathturbation" once you know the win/loss ratios and sizes? Doesn't it all boil down to the probabilities of having say 20 losers straight, and if so, how do you calculate the risk of ruin mathematically, not historically?
I offer these questions in hopes that I might get some clarity and that it might help everyone else who reads and shares on this site.
Conceptualizing ideal bet size
There is more data available for study, than merely your own backtesting results. You could also study the (real-life, real-money) track records of professional traders who use mechanical systems, for further insight about the randomness or non-randomness of trading performance. (Don't forget to add back the management fees and profit allocations, so called 2-and-20, that have been deducted from the published track records of the pros.)
For futures trading this is particularly easy, just sign up to the (free) websites www.iasg.com and/or www.autumngold.com and get access to the trading performance track records of hundreds and hundreds of system trading pros: CTA's and commodity fund managers. Among others, the track records of several Turtles and one Turtle Instructor, are available.
For futures trading this is particularly easy, just sign up to the (free) websites www.iasg.com and/or www.autumngold.com and get access to the trading performance track records of hundreds and hundreds of system trading pros: CTA's and commodity fund managers. Among others, the track records of several Turtles and one Turtle Instructor, are available.
Hi Dallas,
> Wondering if anyone can explain apparently wide fluctuations in test results ...
> ... know the win/loss ratios ...
It's not a simple ratio, independent of anything, but rather a distribution in a high-dimensional parameter space. You mention some of the parameters yourself in the rest of the sentence. Among them are the current market conditions, and at least those change over time.
So you have a time-dependent, multi-dimensional probability distribution, from which you are trying to cut a piece with positive expectation value. With EOD there is not nearly enough data to get a good picture of the distribution. Correspondingly the deviations from the mean expectation values of systems are large. This is another name for "wide fluctuations".
Regards,
Asamat
> Wondering if anyone can explain apparently wide fluctuations in test results ...
> ... know the win/loss ratios ...
It's not a simple ratio, independent of anything, but rather a distribution in a high-dimensional parameter space. You mention some of the parameters yourself in the rest of the sentence. Among them are the current market conditions, and at least those change over time.
So you have a time-dependent, multi-dimensional probability distribution, from which you are trying to cut a piece with positive expectation value. With EOD there is not nearly enough data to get a good picture of the distribution. Correspondingly the deviations from the mean expectation values of systems are large. This is another name for "wide fluctuations".
Regards,
Asamat
Don't underestimate the power of randomness.
Perform the following Monte Carlo in your spreadsheet of choice. Pick an odds of winning (P), a bet size (R), and a payoff ratio of wins to losses (WL). Maybe P = 35%, R = 1%, and WL = 2:1. This is a positive expectancy system, winning 2% on wins and losing 1% on losses.
Now cycle this Monte Carlo through 45 columns or so of 1,000 trades or so each, using the random number generator in your spreadsheet to determine each win or loss. Compare the average equity gains across each column. Note the deviations from the mathematically expected amounts.
Now calculate the maximum drawdown from each column, and compare. Interesting, no?
Perform the following Monte Carlo in your spreadsheet of choice. Pick an odds of winning (P), a bet size (R), and a payoff ratio of wins to losses (WL). Maybe P = 35%, R = 1%, and WL = 2:1. This is a positive expectancy system, winning 2% on wins and losing 1% on losses.
Now cycle this Monte Carlo through 45 columns or so of 1,000 trades or so each, using the random number generator in your spreadsheet to determine each win or loss. Compare the average equity gains across each column. Note the deviations from the mathematically expected amounts.
Now calculate the maximum drawdown from each column, and compare. Interesting, no?