Acceptable Max Drawdown

[RFMontraz]

[i]Clustering means the presence of similar and persistent results in consecutive outcomes, if i am not wrong. Look at the DD graph. the first significant clustering period is the period between 30/4/99 to 30/7/99. For three who months or roughly 65 trading days, your DD is both high and relatively stable. This implies that you are either at the wrong side of the market or your positions are more risky than usual for this whole period. Thats not prudent at all. If in this period for ten days your actual results are 10% daily losses you ll see all your money vanished. Not nice at all.[/i]



As mentioned the losses are capped to a (high) max of 5% per day. I'm thinking of halving that figure now (this won't affect my entries/exits/stops anyway). Once I get in a drawdown it does take some time to get out of it, but honestly it's not stable during the 3 months, as it gradually reaches the peak and then disappear. Regarding the clustering of drawdowns I still don't understand what's wrong with it. I'd rather have a lot of similar contained drawdowns and comebacks than a supermooth equity line and then a couple of shocks here and there. (keeping in mind that whenever there's risk there's vol and dds of course).



PS I'm not trying to desperatly defend the strategy, I just want to understand better certain issues.

[jungle]

go back to nonius' statement about risk tolerance.  if you are running your own money, cool, no probs.  or if you have funding from someone who has faith in you.  but you may find money being pulled if your drawdowns are clustered - they will extrapolate.

[RFMontraz]

Ok. If I reduce my max daily potential loss to 2%, my max drawdowns goes down to 8% and the componded annual return is around 20%. Of course this doesn't affect the drawdowns distribution, which stays, for obvious reasons, the same.



Are we still talking about drawdown clustering? Keep in mind that the drawdowns, even if smaller, recurr with the same frequency, so will people still pull their money? Or the problem of drawdown clustering was related to the fact that, apart from the frequency, the original drawdowns were of a certain magnitude?

[Johnny]

[i]Look at the DD graph. the first significant clustering period is the period between 30/4/99 to 30/7/99.[/i]



There's a confusion of terms here. This period is just one drawdown, it's not a "cluster of drawdowns". Look at the equity chart and you'll see that during this period the equity goes down (the drawdown) and then back up again.



Other than this semantic point, the equity chart and the drawdowns look completely normal for a trend-following strategy. Trend-followers always have their biggest drawdowns just after their biggest victories for the obvious reason that they run their biggest position sizes just when the trend changes direction. C'est la vie, I wouldn't worry about it too much. People that invest in these things understand this is how it works.

[goldorak]

This is a point I was wondering about too. It seems some people have a definition of 'max drawdown' as a loss over a fixed time interval. Personnally, I've always followed the following definition.



If tau is the moment of last maximum of equity :

tau=arg max{0<=t<=T} E(t) where E(t) is equity at time t



then my maximum drawdown is given by :

MaxDD = min{tau<=t<=T} E(t)-E(tau)



With this definition, you get a set of drawdowns which can be analyzed.



Is there a 'commonly' accepted definition ?

[opmtrader]

I agree with Johnny. Sometimes its just the breaks. Drawdowns happen. If you start looking for ways to attack those drawdowns you run a big risk of overfitting. Serial correlation of trades often does not persist over an entire series.



Anthis, lets say this is a trend following strategy. How do you suppose we can make the drawdowns go away? You can't force markets worldwide to start to trend. If you think you are going to do some fancy regime change algo that switches between trend and chop, well, good luck with that. I don't see it happening (and trend following systems have some of the most serial correlation of trades out of all strategies).



Don't burn your data overfitting. Just start on a second, as good or better, strategy.

[Anthis]

Johnny, I dont know details about RFM's strategy. Looking only at the DD graph i see a period of low DD (30/7/98 to 30/4/99) where the graph exhibits relatively low DD, with one exception that reaches -8%, then there is a period  of persistently high DD (30/4/99 to 30/7/99), then again low, then high and so on. Albeit i am not sure how RFM estimates his DD I have a hunch that even the use of different metrics such as VaR, Expected Shortfall, Shortfall Probability will show similar and not contradicting results. Clearly the the graph shows two different regimes for this strategy. One of low risk and one of high risk. I think its wise and prudent to identify those regimes and find possible ways if not to take advantage of it, at least to try to cover himself. Am I wrong?

[Johnny]

I'm making a much more simple point. Consider the following fund values on sequential days: 100, 98,96,94,92,90,95,97,101,103, ... this is just one drawdown of 10% (from 100 to 90). It's not a cluster of drawdowns. Similarly for RFM's strategy the period 30/4/99 to 30/7/99 is just one drawdown, not a cluster of different drawdowns. That's all.

[RFMontraz]

[i]Albeit i am not sure how RFM estimates his DD...[/i]



Anthis, there might be a misunderstanding about the use of the term drawdown here as, the way I see it, it can be calculated in one way only.



As Johnny wrote, if I go from 100 to 90 and back to 100, that's one (10%) drawdown. If from there I go straight to 200, then to 160 and back to 200, that's a second (20%) drawdown.

[opmtrader]

I agree that ideally we should do what we can to improve a system.  However I often find that starting on a second idea usually results in a similar improvement with far less risk of overfitting a single system.  That's just my approach.



Here's an example of diversification.  The blue line is the combined performance of about a dozen systems.  These systems on their own have drawdowns between 18% and 6% and Sharpe ratios around 1.3.  When combined though they have a max dd of only 4% and a Sharpe of 2.877.



[img]/beta/User%20Files/5/Combined_System_Analysis_Combined_Chart2.jpg[/img]



This is not the best example as the concepts these systems are based on are semi-related and the systems are cherry picked, albeit from very stable parameter spaces.  If I wanted to improve the results I would have only included those systems most lowly correlated to one another.  Below is a correlation map of the systems which should illustrate what I am saying.



[img]/beta/User%20Files/5/Combined_System_Analysis_Correlation_Comparison_Chart2.jpg[/img]



Historic correlations amongst systems may change, usually at the worst times.  However we do what we can.  That's what makes system development half art and half science IMHO.  I welcome any commentary on this approach.