|
PureBytes Links
Trading Reference Links
|
Hello list, here's a nut to crack.
The problem is that as long-term systems tend to enter trades
relatively infrequently, even using a huge set of historical data, it
is far too easy to curve-fit the input parameters simply because the
number of observations is not sufficient to prevent this. Therefore,
by running hundreds of tests with *random* entries and combining the
results will, in my reasoning, create a more reliable result set based
on which selecting parameters for the system is more robust and
reliable.
My assumption is that for a long-term system, entry procedure plays
fairly little role in terms of overall outcome from any single
long-term trade. The reasoning is that whatever happens over the
course of 1-2-3-4 months is very unlikely to be expressed in terms of
specific, relatively small-scale entry procedure. What the entry
procedure can provide, however, is the small trade risk for proper
position sizing and max compounding, but is a completely another
topic. But no long-term performance prediction can be derived from the
entry parameters.
Now, having said this about entries, I am eliminating entry procedure
completely for the sake of my long-term performance tests and am
basing entries purely on random number: I'm generating a random number
in the range of 0-1 and everything >0.5 is a buy and everything less
than that is a sell. And after the entry, risk-management/profit
trailing will take over.
The specific procedure is that I will add another dummy input to the
system, which is set to be optimized from 1 to 100, incrementing by 1.
The results from each of the 100 runs are written into a text file,
and read into a database. In the database, I can group performance
figures based on various inputs, and will generate something similar
to TS Optimization Report.
The question to the list is: can you gurus comment on my reasoning
about the different performance parameters, which goes like this:
1) % profitable trades - irrelevant. Ideally should be in the 50%
range, but because of max risk stop-losses, it will usually end up in
the 30% range. However, because entries are random, whatever this
number is is irrelevant.
2) Avg win/avg loss ratio - probably very relevant. This measure will
show how good is the risk management/profit trailing procedure.
3) Profit Factor - also relevant as it is highly correlated to the
*relationship* of the above 2 numbers. Even if we say that 1) is
irrelevant, 2) is still important.
4) Net P/L - as it reflects more or less 3), it is equally relevant.
5) Number of trades - irrelevant. It's clear that when, due to the
random luck, more loosing trades are entered which get closed out
quickly, this measure should not be used with random entry testing.
6) Expectancy - relevant, correlates to 3).
Any other suggestions?
Best regards,
Ivo Karindi
|