Way of the turtle

176
•
Way of the Turtle

---

days and then resuming the previous position size when that num-
ber of days had passed. I found two such periods in the 10 years of
data I used for testing. If one lightens up 96 percent for the first 2
days of September and the first 25 days of July, one can get better
results. How much better?
The addition of this rule improves our returns, which move from
45.7 percent without the rule to 58.2 percent with it, and our draw-
down goes up a tiny bit from 39.2 percent to 39.4 percent, whereas
the MAR ratio goes from 1.17 to 1.48. Again we think: “This is a
great rule; my system is now much better.”
Unfortunately, this rule works only because there were significant
drawdowns during those periods in the past, not because there is
something magical about those periods. It is not very likely that the
drawdowns will occur on the same dates in the future. This is an
example of the worst kind of overfitting, but you would be surprised
how many otherwise intelligent people fall prey to this sort of trap.
If you do not know this, you might think that the system is good
enough to start trading with it. You might even start to raise money
from friends and family by telling them about this wonderful sys-
tem and its results. The problem is that you really have a system
that returned 41.4 percent, not 58.2 percent; a drawdown of 56.0
percent, not 39.4 percent; and a MAR ratio of 0.74, not 1.48. You
are bound to be disappointed with the real performance because
you have been seduced by the easy improvements of curve fitting.
Next I’ll examine ways to avoid the problems discussed in this
chapter. I’ll show ways to determine what you actually may be able
to achieve from a system in order to minimize the impact of trader
effects, detect random effects, properly optimize, and avoid over-
fitting to the historical data.
Lies, Damn Lies, and Backtests
•
177

---

This page intentionally left blank 

---