EdgeWare FastBreak Pro Version 6.2 - Manual (Page 83)

83
usually results in a reduction to the variable being optimized (constraints such as MDD,
or switches per year have the same effect). We are willing to pay a small price in IS
performance if it results in a system that has better OS performance, i.e., better predictive
ability. What is probably happening in the above example is that the robustness is
preventing the genetic algorithm from converging too rapidly.
Trying to satisfy the
robustness constraint is similar to the effect of mutation. Preventing too rapid of a
convergence forces the genetic algorithm to do a better job of examining the total trade
space (“trade space” is a way of saying full range of parameters) . This results in finding
better parameter solutions.
In some cases, you may see a significant difference in performance with a small
percentage change in parameters. For example, you may have a small trading family with
only four members. Using a Top% value of 59%, the sell point is reached when a fund
drops out of the top two ranking positions (0.59 X 4 = 2.4 which FastBreak Pro rounds to
2). When the 59% Top% is increased by 10% the sell point requires the fund to drop out
of the top three ranking positions (0.59 X 1.1 X 4 = 2.6 which FastBreak Pro rounds to 3).
You must be aware of these types of effects.
It may be emotionally difficult to select trading systems that use the robustness check and
do not perform as well compared to systems that skip the robustness check. However, the
majority of trading systems developers would argue that systems sensitive to small
parameter changes should not be trusted.
Issues to Consider when using the Trendline Options
Minimizing Runtime when using the Trendline Options
As noted earlier, the trendline algorithm is very computer time intensive. Because it
takes so much time to calculate each trendline for each member of the trading family
FastBreak Pro will calculate all the versions of the trendline for each fund just once. This
is to say that if you give trendline Buy Filter a “Size” range from 0 to 10 there are 11
trendlines that need to be calculated for each member of the trading family. If your
trading family has 100 members that requires calculating and storing 1100 trendlines.
FastBreak stores these different versions in an array and uses them over and over as the
genetic algorithm chooses them.
The FastBreak Pro optimizer builds this array at the start of the optimization process. It
can take an hour or more to build all these trendlines for a large family. The good news is
that once all the trendlines are built and stored the remainder of the optimization process
will run at normal speed.
NOTE: When you start the optimization process you will think the program is locked
up. We highly recommend that when you first experiment with the trendline option to
use very modest family sizes.