EdgeWare FastBreak Pro Version 6.2 - Manual (Page 80)

80
Adjusted Performance = 20% x 15/18 x 10/14 = 11.9%
This calculation is made for all systems (chromosomes) in a generation. All the adjusted
performance values are put in order, from best to worst. At this point, only the best
systems are allowed to “survive.” The percentage to survive is controlled by the user, but
a good starting point is 50%.
The next step is to combine system parameters to produce hybrid offspring. This can be
thought of as taking some aspects of one system and combining it with aspects of another
system. We will not go into the mathematical details, but the idea is to try to find better
combinations of trading options and parameters. The result of this combination is an
“offspring” or “child.”
The offspring may also be “mutated” using a random process.
The effect of mutation may effect a positive change or a negative change. You will not
know until the new trading system is evaluated. Reasonable mutation rates are 2% to
10%. If you see that your runs are not improving from one generation to the next, this
can be due to premature convergence on a local optimum (as opposed to a global
optimum), or to the fact that the mutation rate is too high which is providing too much
randomness in the process. You can try increasing and reducing the mutation percentage
to improve the optimization.
The offspring form the members of the next population and the process begins again.
Improving Robustness
One of the most serious concerns with trading system development is the problem of
system “robustness.” Robustness is a measure of how a trading system will perform
under a variety of market conditions. The goal of any trading system is to perform well
under a wide variety of market conditions.
Trading systems that have been over-
optimized very often are not robust.
One indication that a system may be robust is if small changes in trading parameters do
not seriously affect historical performance. For example, if using a moving average stop
of 50 days works well and using 45 or 55 days also works nearly as well, then this could
be considered robust.
With the standard version of FastBreak, we recommend using the contour graphing
capability of FastGraph to examine how parameter changes affect performance. While
this is very effective, only one or two parameters can be examined at a time. We have
taken an approach with FastBreak Pro that is similar to what a user looks for with
FastGraph, but we have incorporated it in a manner that allows the software to make the
evaluations rather than relying on human interaction.
FastBreak Pro performs an automated sensitivity analysis of parameters. After FastBreak
Pro has evaluated the trading system, two additional runs of the same system are made,
but with a small variation in all the parameters. First, the parameters are increased by a
user-defined percentage using the following option on the GA Screen: