|
PureBytes Links
Trading Reference Links
|
you don't need any historical data to test this type of system.
you can create synthetic price time series with various
key statistical characteristics and test the system on that.
there are several variables that you can play with there.
trendiness, noise, volatility, etc...
you can recreate virtually any price time curve and study
how your system works on that curve.
there are about 4 "modes" that any market is going through
more or less:
- trend
- countertrend
- sideways tradable
- sideways untradable ( dead, flatlining markets )
so with synthetic time series you can easily model any of those types.
but there are no perfect boundaries between the modes so
the best way is to come up with an algorithm that does not
differentiate between those market modes.
you don't really need to id those modes and then switch the trading
logic. because it will get hairy on you real fast...
to clean things up mathematically
the proper algorithm should not see trend, countertrend or
sideways markets. that's the hard part, trying to come up
with the solution like that. and the only way it could be accomplished
is if the algorithm does direct or indirect predictions.
another thing is that if the system is adaptive you don't need to backtest
it.
you throw it on any price curve and all you need is a ramp up period
during which the system picks up on all major statistics of the price curve
and then makes bets. the ramp up in most cases needs not to be
too long. a thousand data points is more than plenty. but more like
100 points is all you need for the adaptive system to capture the current
situation.
another point is that you have to design the system logic to be
separated into two algorithms:
- one should handle winning trades only ( predictions that worked )
- the second one should handle losing trades only ( those that failed )
and the second one is a little bit more important than the first one.
it's a very important key.
and as i have mentioned before since the statistical footprint of
the market the system trades will be gradually changing as markets
evolve. the system, if adaptive, will pick that up. but the logic of
the system should applicable to all price curves and should not
be affected by the changing market footprint.
as far as the depth of testing. forget about it.
if you have just a system and you think that by testing it on longer
periods is the correct way of doing it, it's not so.
all you will do is overgeneralize the system and the result will be
marginal. if you want to tweak the system you need to pick
proper time period. not too short but not too long either.
if you pick too short you will curve fit for sure and if you pick too
long you will overgeneralize for sure.
the best way to pick the period is to mimic the proper system logic.
- find the period of the price curve where the statistical price
characteristics ( behavior ) is constant or close to constant
or what is called close to being stationary.
- optimize over that period
- cross your fingers and hope that it will continue to be
within those stat boundaries for the trading period.
- as soon as it changes ( you have to eyeball it ), reoptimize.
example:
take, say whatever, ok CSCO. for the past year or two
the stock went from 100 tabletop to 50 table top, to
30 tabletop to 20 table top where it is trading now.
the stock statistical characteristics have changed completely,
the liquidity levels, the trendiness, the volatility levels
everything changed since it was under 100 dollar stock up
to 15 dollar stock. so if you try to run your intraday system
all the way from when csco was a 70-80 dollar stock till
now, it's not going to work...
so the proper way is to tweak as the stocks moves from
one tabletop to the other.
same goes about the SP. when sp was trading at 500 level
is not the same anymore as it's trading now at 1000 level or
1500 level. the statistics change when those tabletop change.
take another example with bonds. when bonds traded at
the old pit, the pit was pretty tight and we had days when
bonds flew a point or a point and a half in under 30 seconds
on those reports.
now that the pit is huge and bonds are losing their
importance and volume that ain't happening much anymore.
so when a contract is changed or the system is changed
statistical characteristics of price curve change with it.
that's why it's important to know when to tweak and
how much data to use to tweak.
but again an adaptive system should not need too much
data to pick up on those changes...
to stress test the system you have to drop it on synthetic
price curve and model rapid or slow footprint changes and see how
the system reacts to those changes. then you can see
where the weaknesses are.
bilo.
----- Original Message -----
From: "Brian" <blink64@xxxxxxxx>
To: "List, Omega" <omega-list@xxxxxxxxxx>
Sent: Sunday, April 29, 2001 6:54 PM
Subject: RE: Trading Systems
"here is an easy rule on how to
> id a good system. if you can load your system
> in the chart and follow ten signals in a row without
> any hesitation, regret, worry or low confidence than
> you system is great. if however you question the
> signals, you double guess, you hesitate and you think
> twice that means your system is no good...
Some good points made here. But what gets you to this point isn't just how
much money your systems makes but also the depth of your historical testing.
You need a good system to trade confidently and a good system is defined not
just how much it makes but also by how it performs over different market
modes (up, sideways, down). How many systems have had great runs just to
give it all back and then some? I usually need no less than a year of real
time data (or at least a market with 2 obvious modes but I strive for 3) to
get confidence to do what bilo suggests defines a good system (on 60 min
bars). The shorter the time frame your system is, the smaller amount of
time you will probably need because you will get all 3 market modes in a
shorter period of time. This is one limitation of TS Pro. The 6 months of
real time data doesn't offer enough backtesting for me to trade confidently.
For me, the amount of money the system makes isn't the only factor, also
important is how many trades over what period of time and markets I've
tested against. When my system performs steadily in what I think are
different market modes then I can trade it without question because I know
that the odds are good, whatever the system loses it will most likely at
some point get it back. It's just a matter of reliably placing the trades.
But I need sufficient and good data to get to this point.
Brian
|