Forums  > Trading  > Max DD as a risk measure  
     
Page 1 of 1
Display using:  

Energetic
Forum Captain

Total Posts: 1423
Joined: Jun 2004
 
Posted: 2018-01-08 18:22
I've been surprised as to how much attention people pay to Max DD.

Let's compare two imaginary strategies with equal returns and Sharpes but different DDs. One drops down -50% about once a year and recovers in a week, on average. The other dips -25% every three months and recovers in a month.

Is it reasonable to consider first one riskier than the second or the other way? My instincts are going with the second. If I had money to invest I'd be looking at something like an integral of DD while under water between two high watermarks.

Because DD essentially represents a loss that you'd take if you have to sell at a bad time, correct? (Personally, I disagree b/c it's not really a loss but a paper loss vs. the max that one could realize if sold at the high - kind of nonsensical view, no?) But if max DD is associated with a short narrow trough then your chance of being forced to sell during that particular week is very small. So why worry so much about Max DD instead of more meaningful statistics, like what I suggested above?

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

AndyM


Total Posts: 2322
Joined: Mar 2004
 
Posted: 2018-01-08 19:24
If you consider both risk of ruin (even in a world of zero uncertainty) and the impact of Knightian uncertainty, then max DD seems highly pertinent. This is even before you add in psychological factors. You will have many more takers for the second strategy than the first because it affords you the luxury of being wrong; the first doesn't.

I used to be disgusted; now I try to be amused...

Patrik
Founding Member

Total Posts: 1339
Joined: Mar 2004
 
Posted: 2018-01-08 20:30
b/c it's not really a loss but a paper loss

Unless one is talking about accounting rules vs "reality", the term paper loss is a bit of a slippery slope. And often signals lots of overconfidence :)

I guess I'm not in love with max DD as a measure, but it or something like it gives some useful rough number to signal when something may have changed and you may need/want to question your assumptions. I.e. if something has been running for 10y with say 10 trades a day and you are starting to approach a DD that is 2x anything you previously experienced - it's natural to question if something has changed. The outcome doesn't have to be that it has, but it at least makes sense to take a look in the mirror. Perhaps you conclude the world has changed and you should throw in the towel (reduce risk allocation, kill it, whatever). To make those decisions relative to a lower max DD is less scary than vs a massive expected DD. Extreme example: for a strategy with an expected DD of 95% once every 5y with a recovery of 1y, if you start to see something worse than that, it's a bad situation to wait for and be in before throwing in the towel..

(obviously these things doesn't have to sit outside the strategy/model, you may have a more adaptive risk allocation process continuously calibrating vs history etc - what I'm getting at is more the intuition as I see it)

Capital Structure Demolition LLC Radiation

Energetic
Forum Captain

Total Posts: 1423
Joined: Jun 2004
 
Posted: 2018-01-08 20:52
@Patrik

Paper loss was a bad choice of words. Perhaps, I should have said - imaginary loss.

I'm not at all denying the usefulness of DD as a quantity monitor so I agree the whole paragraph wholesale. I was thinking in terms of strategy evaluation after backtesting stage - deciding whether to invest in it.

@Andy

From what I hear, you're right: more people will prefer second strategy. I just disagree that this is rational.

If you're a hedge fund with a portfolio of strategies then you risk losing no more than what you allocated to this one. Not that it would be pleasant but it won't be a real ruin. At least one should consider max DD in the context of its duration and relative to profitability. That's obviously just my uninformed opinion.

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

darkmatters


Total Posts: 70
Joined: Nov 2010
 
Posted: 2018-01-08 23:50
Two points:

1) Max drawdown is like sampling from the tail of a distribution, and in finance a fat-tailed distribution. Even assuming stationarity, the true 'max drawdown' of a PnL distribution is always (much) larger than what has been seen before. That is why it is rational to not prefer that large drawdowns.

2) Max drawdown numbers are very prone to overfitting. Lets say you have a strategy that has some 50% loss in a day or two. There is always some seemingly simple rule that could have you reverse position the day before the big loss that would erase that drawdown. The potential investor would never know you applied it post-hoc either.

NeroTulip


Total Posts: 999
Joined: May 2004
 
Posted: 2018-01-09 09:14
Ha, if only MaxDD was a robust statistic, and not highly dependant on the sample...

"Earth: some bacteria and basic life forms, no sign of intelligent life" (Message from a type III civilization probe sent to the solar system circa 2016)

ronin


Total Posts: 221
Joined: May 2006
 
Posted: 2018-01-09 13:21
Energetic,

Say I give you a strategy that makes $1 per day, and loses $ 1 mln every 100,000 days. That is once every 400 years. You backtest it over a ridiculously long time period - say all the way back to 1980s - and you see Sharpe infinity. The strategy is of course negative.

I am exagerating, but only slightly. Have a look at last year's Euromoney hedge fund awards and tell me how much I am exaggerating.

That's the problem in a nuthsell. The notion of "return per unit of risk" is sound. The issue is what the "unit of risk" is.

Standard deviation makes sense as a unit of risk - *if* your return distribution is Gaussian. If you are selling protection like my $1 up - $1 mln down example, it is completely meaningless.

If you are selling protection, then of course maxDD is perfectly meaningful and standard deviation is meaningless. Theoretical maxDD, historical maxDD, simulated maxDD etc - they all have their place.



"People say nothing's impossible, but I do nothing every day" --Winnie The Pooh

Energetic
Forum Captain

Total Posts: 1423
Joined: Jun 2004
 
Posted: 2018-01-09 17:22
No objections to anything stated here. I didn't mean that max DD can be ignored and I certainly didn't mean that large DD is good. I just don't think that it is necessarily the most interesting statistic (although in some cases it might be).

However I believe an integral measure will be more robust, that's exactly the motivation. In particular, it could partially address the overfitting problem mentioned by @darkmatters.

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

chiral3
Founding Member

Total Posts: 5030
Joined: Mar 2004
 
Posted: 2018-01-09 20:11
The other thing I will add is that, from a historical perspective, as people searched for ways to evaluate hedge fund performance through the 00's, which was really tricky on both an absolute and relative basis, maxDD became a more important tail risk / ruin metric (survivorship bias notwithstanding). Comparable or coherent risk measures weren't really available or realistic and maxDD is a very simple stat to compute and communicate.

Nonius is Satoshi Nakamoto. 物の哀れ

Energetic
Forum Captain

Total Posts: 1423
Joined: Jun 2004
 
Posted: 2018-01-09 21:01
See my sig line ;)

For every complex problem there is an answer that is clear, simple and wrong. - H. L. Mencken

chiral3
Founding Member

Total Posts: 5030
Joined: Mar 2004
 
Posted: 2018-01-09 21:07
Off topic, but remember in the thousands how managers called alpha "skill"? As in "There's 200bps of skill." A recent article, I think it was in the WSJ this past weekend, said that if a strat was employed that (re)invested in the top performing funds ('40 act) of each year since 1991 the return today would be -18% pa.

Nonius is Satoshi Nakamoto. 物の哀れ
Previous Thread :: Next Thread 
Page 1 of 1