PureBytes Links
Trading Reference Links
|
Gary.
I think the Walk-Forward-Optimization should give you a more realistic
expectancy for the unknown future (you know where to get samples)!
On the drawdown issue, your saying about the %DD is only true if you
test one market at the time. If you test multiple markets and use a
money management method that is based on the Overall equity then your
method would NOT give you a valid figure. Also depending on what MM
method you use on the portfolio level, the starting date can have a
major impact.
Regards.
Volker Knapp
Wealth-Lab Inc.
www.wealth-lab.de
www.wealth-lab.com
-----Ursprüngliche Nachricht-----
Von: Gary Fritz [mailto:fritz@xxxxxxxx]
Gesendet: Wednesday, April 30, 2003 7:03 PM
An: omega-list@xxxxxxxxxx
Betreff: Re: when a strategy breaks the max drawdown
> If you measure the DD in % and you design a robust trading strategy
> you maximum drawdown simply shouldn´t be broken during the first
> time of real time trading.
Sorry, I can't agree with that. Measuring DD in % is valuable,
but it's not a magic panacea. Stridsman claims the TS stats are
useless because TS shows DD in $ instead of %, but all you have
to do is measure your DD in $ for a fixed-position-size system.
The example of a $10 DD with a $100 stock vs. a $200 stock is
flawed, since it's comparing the $10 DD with different position
sizes. With a fixed position size, the system would buy 2x more
shares at $100 than at $200, and you would see a 10% DD in the
*position* for a $10 DD off $100 or a $20 DD off $200, just like
you should. With fixed position sizes, the dollar-denominated
stats in TradeStation are just as valid as percentage-denominated
stats, in spite of Stridsman's claims. You just have to divide
the $ DD into the $ position size to get the %.
And neither $ or % DD guarantee you won't exceed the level of DD
you saw in your test. Unless you have some amazing way of
developing a "robust" trading strategy, I think the market will
still find ways to pop your bubble.
Example: I have a simple system that trades the ND. With
volatility-scaled position sizes, its equity curve is nearly
straight all the way back to the inception of the ND in 1996,
even though I only tuned it on 6-12 months of data. I would
consider that to be reasonably robust. But back in late 2001
something changed in the ND market, and in 11/01 that system just
plain stopped working. It went into an unprecedented drawdown
and has never recovered to this day. No amount of retuning
allowed it to work in the post-11/01 market conditions.
Fortunately I discovered a change in bar size allowed the system
to react quicker to the market. That new configuration seemed
much more robust, and it performed well from 1996 through 2002.
And even that setup is stumbling in the recent contracted market
conditions.
I can find minor tweaks that allow the system to handle recent
conditions AND also handle all the wildly varying conditions
we've seen over the past 7 years, even though 80% or more of the
7-year test is out-of-sample. And still the market manages to
find new ways to hoodwink me, and I have to find new tweaks that
include those new conditions.
If you have some way of verifying that a system is "robust"
enough to handle the unforeseen changes in the future, I'd like
to hear more about it. I don't remember Stridsman having any
remarkable methods beyond his %-based approaches.
Frank Fleischer wrote:
> Max DD is a deceiving and worthless piece of poop. It's deceiving
> because traders think that it's a valuable statistic. It's
> worthless because it says nothing about anything since any Maximum
> implies a single unusual event that took place within a finite set of
> historical data. Single historical events say nothing about the
> future.
Exactly! And remember that when you optimized your system, you
carefully (even if unknowingly) filtered out any settings that
had bad drawdowns IN THE PAST. That doesn't mean you filtered
out all settings that CAN have bad drawdowns IN THE FUTURE. It's
almost certain your system still has the potential to generate
large drawdowns, even if it didn't in your backtest.
Gary
|