I recently joined C2 and listed some systems. My fist suggestion is to replace “annualized return” with something more meaningful. Look at this summary page for my systems:
http://i34.tinypic.com/qz57ie.png
The “annual return” is so enormous that it doesn’t even fit the line. This is obviously nonsensical. I did notice that there is another set of performance stats, so perhaps one of those should replace it.
My second suggestion is to add another performance metric, which is defined as:
SQRT(N) * (NetProfit / N) / STDEV(P&L)
Where N is the total number of trades
NetProfit is the profit after N trades
STDEV(P&L) is the standard deviation of profits and losses over the entire period
My comments: The new metric you propose is called the "System Quality Number" by Van Tharp, although I have also seen it in other books (with no name for it given). It is a t-test type metric, that is supposed to give you the confidence that a greater than zero avg trade result is statistically significant.
Major drawback is that it really should be standardized for N, say the number of trades in a year, or by using N=100. Otherwise, I could have a pretty low avg/stdev with N=10000 and the overall number would be great (the system on the other hand would not be).
This metric would take quite a bit of explaining to be meaningful, which likely means people will unfortunately ignore it. I do use it in evaluation of my own systems though.
Cool metric, I’ve never seen that before.
Your system is too new to have a normal annual return. Wait a couple of days and the number will fall into a normal range.
Looks like you hit number 9 black after gambling and losing 47% about.
I agree that it should be normalized for the the number of trades in a year (or some other fixed period of time). I use it for ranking my systems when backtesting and optimizing, which is always over the same period of time for all systems, so normalization is not needed. For C2 purposes, yes, a normalized version would allow to compare different systems of various life spans.
SQRT(N) * (NetProfit / N) / STDEV(P&L)
This is not relevant to hardly any of the systems here. I know the theory, but I don’t think it matters. People will be able to realize annualizing beyond 30 quintillion % is probably not sustainable, or even recommended to try to keep up with.
I don’t think we need a special statistic for your system.
I’ll tell you the Sharpe ratio is much more explanatory than this.
I’m curious about your statement "This is not relevant to hardly any of the systems here. I know the theory, but I don’t think it matters."
Can you explain your thinking?
Thanks
The Sharpe ratio is better for these situations.
That is correct, but that’s exactly my point. No sane trader will use my system, yet it’s currently the top performing system on C2, based on the annual return. Annual return is meaningless by itself, so why show it?
OK, maybe it is just me, but now can you explain “these situations?” Are you talking about his system, new systems in general, or what? I am confused and I’m trying to understand.
There is no better formula that captures risk and reward better than the Sharpe Ratio. So much so, that Mr. Sharpe received a nobel prize. It doesn’t get any better.
(APR-rfr)/stdev of returns. Simple and very useful. A normalized risk adjusted ranking. C2 emphasizes this a lot. Number of trades has no bearing on this statistic.
Eugene, every system that starts out doubling in its first few days will have huge APRS. When your system is around for 9 months, you’ll see the real statistics, and I’m sure you know that. We just had the MMO system have its 9 month stint and crash. Whatever you’re doing isn’t what I’d consider a holy grail, by any stretch of the imagination. It’s very presumptuous of you to call it that, when it’s so far from being that.
You’re not the first vendor to knock it out of the park out of the gate. For some on here, it’s actually a system rule that if you don’t boom it out of the gate, then kill the system and start over.
It’s a scalping system, and not a very good one at that. A 0 APD means exactly 0 profit, so it wouldn’t even be possible to have that combination.
It’s a scalping system, and not a very good one at that. A 0 APD means exactly 0 profit, so it wouldn’t even be possible to have that combination.
It’s APD would be something like 0.00001, rounded to 0. It has 1415 trades in 3 months, a scalping system.
I understand that. My point is, however, is that there are better performance metrics which are less sensitive to (or even independent of) the trading system’s age. As it stands right now, my system’s APR as reported by C2 is bigger then the number of stars in the observable Universe, which is in the order of 10^22. That’s a throwaway piece of trading statistics.
I don’t believe it’s needed, Eugene. It’ll just complicate an already complicated set of system statistics. Once you’re around for a few more months, we’ll see the real stats. Nobody pays attention to stats until at least 3 months out, imho.
At first glance it would seem wiser not to report any statistics based on such small amount of data. However, if this is done then every week some new vendor will ask why his system has no statistics.
That the APR has a ridiculous value in young systems does not imply that it is generally a meaningless statistic. It is just extremely unreliable if it is based on a small sample. All statistics are unreliable if based on a small sample. This does not say anything about their meaning in the long run.
E.g., try to compute the t-statistic that you suggested when you make 4 trades with $500 profit in each case.
Nevertheless, I think it is an interesting suggestion. Note that the multiplication by sqrt(N) makes the statistic very sensitive to the total number of trades - in the same way as the ordinary t-statistic is sensitive to the sample size. To some extent that is an advantage, because it will partially suppress the effect of random luck in young systems. However, this also obscures some of the relations between systems: A very effective system may get a lower score than some mediocre system just because it has a smaller number of trades. So I would also like to see the effect size part of this statistic (i.e. without the factor sqrt(N)).
Furthermore it seems to me that the statistics is vulnerable for some manipulations. When a vendor is in a winning trade he can close and re-open the trade a few times just to get a higher N and lower SD. Then his t-value will probably go through the roof.
(continuation of my previous post)
"My point is, however, is that there are better performance metrics which are less sensitive to (or even independent of) the trading system’s age. "
Such statistics are generally impossible. In order to get the variance independent of N, such a statistic must contain a factor that depends on N (usually sqrt(N)). But then the expected value will depend on N, unless the system has expectation 0.
