EdgeWare FastBreak Standard Version 5 - Manual (Page 64)

64
The difference between the curve fit lines is 2-5 % for the better performing test cases.
The trader would need to determine if 2-3% per year of performance is worth reducing
maximum draw down by 2-5%
Finally, what are the best parameters to use for buy and sell ranking? The values that
gave the best annual return may not be the best choice because that combination may
have bad characteristics, i.e., high draw down, too many switches per year, etc. The most
significant problem - the parameters may not be a good “general solution” and only a
“statistical fluke.” Stability is a requirement in your parameter choice. In other words, if
the parameters are changed slightly, there shouldn’t be a large change in annual trading
performance. You do not want to “curve fit” history.
Here is one method of selecting potentially good parameters.
Go into your spreadsheet that has been previously reordered in decreasing annual return
order. Next, make an X,Y scatter plot using the buy ranking period as the X parameter
and the sell ranking period as the Y parameter. Plot only the top ten best test cases.
What you will see is the following scatter chart:
10 Best Test Cases
0
5
10
15
20
25
30
35
40
45
50
0 10 20 30 40 50
Buy Ranking Period, days
Sell Ranking Period, days
2
8 3
4
6
1
9
5
10
7
The data points are numbered from 1-10, where 1 is the point with the best performance.
In this chart, there is a cluster of points around the 23 day ranking period. This cluster
implies that annual return does not have a large variation in this region of ranking pa-
rameters. There are other parameters, 42 and 46 days for example, that are isolated and
probably should be avoided. It is better to select a pair of buy and sell ranking periods
that have slightly less performance but provide a more general solution. Sometimes there
is the best of all worlds-best annual return and good parameter performance stability.
The sell ranking period and buy ranking period do not need to be set equal to each other.
If a large number of variations on buy and sell ranking parameters (39 buy ranking peri-