|
PureBytes Links
Trading Reference Links
|
Michael Stewart wrote:
>
> Hello Frank
>
> JJ> Hey, why not use what works now, and stop worrying about whether
> JJ> or not it backtests successfully for several years !
>
> Because markets evolve. If I remember rightly Oddball required a 3-1
> ratio of adv:dec to make an entry 'worth while'. The same ratio of
> 3-1 is used by the military to just about 'ensure' victory over the
> oppo.
OddBall is using Advancing Issues only, Declining issues are not used at
all.
> If however you work off 1-1 and volatility drops which it
> certainly will (as sure as night follows day) you will be trading a
> smaller range in less selective periods. Double whammy. Of course you
> could add a filter - DH's (brilliant) noise filter for instance or a
> simple minimum movement or range filter or atr as a condition such as
> Bob Fulks suggested. Which is generally why it makes sense to ensure
> your system performs in most market scenarios for 10 years or so,
> admittedly since 9/11 we've seen both up and down but largely BIG
> daily trading ranges.
I've had "better" results using
1. the difference between Advancers ROC and Decliners ROC, still using
1-day RL lookback
2. more smoothing (and higher ~2hr lag) using a longer period
moving-average
3. higher threshold levels to buy/sell (+20 / -20)
4. Gap filters
Your point about a method that holds over time is very important. In my
testing, most of the filters I tried (Adv/Dec Vol, Net Adv-Dec >0, TICK,
fewer hours e.g. 1100-1530 EST etc) didn't stand the test of time (still
just 3 years of testing). I tested down to 5min bars. Same goes for all
the obvious price direction confirmation signals. Gap filters (make it
easier to buy after a gap down and vv) did yield some improvement.
Ofcourse, if I let the TS optimiser work on it, to "tune" the lookback
periods, threshold levels, AMA period etc, it'll results will look
astounding.
The basic version I described, w/ gap filter, is producing NetProf 470k
with MaxIDD < 40k over 400 trades in 3 years (see attached GIF). It
still isn't adaptive and it still makes no use of Data1 (price). Yet I'm
concerned a the low MaxIDD I achieved is due to curve-fitting.
Regards, Michael
Attachment:
Description: ""
|