|
PureBytes Links
Trading Reference Links
|
some thoughts on FL and NN,
the results i have seen from fuzzy pete's postings based on
his software have a few important characteristics,
1. win/loss % about 50/50, most likely < 50/50
2. the $ gain on average > 0, imply supposed positive return
3. all european mkts
I have used a lot of NN / FL for research assistance ..
so I think I understand the value of them ...
My view is that implementing a trading system based on FL or NN
with results like #1 and #2 will lead to bankrupcy.
#1 - a great example of systems with less then 50% bias on price
distribution is something like flipping a fair coin. since a fair coin
gives 1/2 chance to win, and the distribution of a long sequence of
coin flip (heads and tails) will tell you 6 to 7 continuous heads
is pretty slim in chance. so if you bet consistently on tails after
5 to 6 consecutive heads should give you a better chance to win,
mathematically. you can even double down. The problem is the
gain is always a low fraction of the total sum of the bet size, and
the expected return is still a negative number. In long run, you
will lose it all after a sequence of failures hit.
#2 - $ gain averages or % gain averages are not dependable
measures of a good system. It just show how the system's been
curve fit (all backtesting are some form of curve fit) to the historical
data without appropriate understanding of the properties of the
price stream to be traded. Thus trading system with just $ gain
or % gain numbers looking ok is committing suicide.
#3 - european mkt - mkts dying, no real vol, and not mature.
thus, they "trends" and no need of any "rocket science" to trade.
breakout will do the job 90% of the time. THIS is understanding
of the mkt by studying their properties first.
DID I mention shorting euro since its existence :)
US stock mkts are more mature as the participants in general has more
access to trading knowledge and in general higher risk takers.
Thus the mkts are more volatile in general. pure breakout will
work in only a particular segment of these mkts.
FL and NN after automatic training and optimization are just
million of signals and millions of rules combined => each signal
are not well understand and each signal could be bad, but cover
up by the summarized result of profitability as a whole.
Thus if there is one rule in there that is formed but not used
historically - that is the rule which probably will break
ones account.
As a summary, the problem with FL and NN is generalization.
And the concept of appropriate training of them has no
standard to measure against.
They are the best tool for finding direction in reseach though,
if they pick up something unusual and that is something you never
thought of.
-Lawrence Chan
|