Adaptive Trading Strategies

Quote from Muskoka Joe:

Another thing to consider is that it is not the the math you are throwing at the premise, but what "premise" you are throwing the math at.

Moving Averages are arbitrary with no pattern of distribution of interaction that will be consistent. While past price patterns and derivations thereof can forecast the probability of future price being higher or lower, just not in a MA crossover.

Finding a solid premise is the trick, not trying to slay a useless one with intellect. All the math may possibly help a genuine edge be better.
More...
The problem with this line of reasoning is that you lump all moving averages together when they really have very different efficacies. I know almost everybody here has heard of adaptive moving averages but most have not gotten their hands dirty, so to speak, by actually diving into the inner workings. Most have tried a canned program (usually Kaufman's moving average, which by the way I can assure you is programmed inaccurately more often than not) and concluded that AMAs were no better than conventional moving averages. While I do agree that a MA crossover is hardly a sophisticated trading system, it gives some semblance of the potential of finding the best moving averages and using them along with other financial indicators in a well-designed trading strategy.

Regarding simple moving averages, I once said that expecting a SMA to remove the noise from your data is like expecting a meat grinder to remove the fat from your steak. I stand by that assessment, which is why I have concentrated on AMAs.
 
Quote from kut2k2:

I was challenged to demonstrate the efficacy of my own adaptive moving average so here are my test results.
More...

a ma system, faded, and on range bars, can do very well.
 
It is not the type of moving average that I was speaking to, it is that the market does not hide its edges dependant on the optimized lookback period of smoothed data --- no matter the smoothing method. Any method that smooths data and then looks for a crossover to signal an entry in the traditional sense won't provide a sustainable edge you would really want to trade. In trending markets its very blunt approach will result in profits, but the drawdown wouldn't be tolerable for most, and much better options exist.

Re post above, non traditional thinking starts to get you somewhere. Once you can get a thread, and pull on it, you will start to open up a lot of doors, you would only arrive at by being immeresed in the process. The stuff that works all the math geniuses in the world could not think of without some market time IMO.

Quote from kut2k2:

The problem with this line of reasoning is that you lump all moving averages together when they really have very different efficacies. I know almost everybody here has heard of adaptive moving averages but most have not gotten their hands dirty, so to speak, by actually diving into the inner workings. Most have tried a canned program (usually Kaufman's moving average, which by the way I can assure you is programmed inaccurately more often than not) and concluded that AMAs were no better than conventional moving averages. While I do agree that a MA crossover is hardly a sophisticated trading system, it gives some semblance of the potential of finding the best moving averages and using them along with other financial indicators in a well-designed trading strategy.

Regarding simple moving averages, I once said that expecting a SMA to remove the noise from your data is like expecting a meat grinder to remove the fat from your steak. I stand by that assessment, which is why I have concentrated on AMAs.
More...
 
Quote from Muskoka Joe:

It is not the type of moving average that I was speaking to, it is that the market does not hide its edges dependant on the optimized lookback period of smoothed data --- no matter the smoothing method. Any method that smooths data and then looks for a crossover to signal an entry in the traditional sense won't provide a sustainable edge you would really want to trade. In trending markets its very blunt approach will result in profits, but the drawdown wouldn't be tolerable for most, and much better options exist.

Re post above, non traditional thinking starts to get you somewhere. Once you can get a thread, and pull on it, you will start to open up a lot of doors, you would only arrive at by being immeresed in the process. The stuff that works all the math geniuses in the world could not think of without some market time IMO.
More...

Goog post. Using tools in ways others do not can open doors.
 
Quote from flyingdutchmen:

fwiw, i still do not agree with the name you have given the aproach, but the aproach itself i find a wonderfull thing if the calculating does not take to much time. i havent seen any retail platform with something similar yet, if i would want to do this i would need to optimize by hand every new bar to find out the suitable sets of parameter which would have givin me the greatest profit over xx bars for the upcoming trade. i am sure the average retail trader would make great use of it if a platform like tradestation would offer a similar aproach. for this you have my respect
More...

I'm commenting on this thread late and so apologize if it has moved on from when this comment was made. But since I can be regarded as falling into the expert category on how this software works I can add some useful comments on the adpatation/optimization process.

What I think people are assuming, and which hasn't been explained properly, is that a best (optimal) set of numbers is ferreted out of the data on every bar after a termination criteria is met. And, that either the best or average of those "bests" is used for signal generation. This is not correct in this instance. By default, Dakota allows swarm agents to move once per bar relative to the performance landscape presented to one another. I am unaware of any optimization scheme that provides an optimal result based on a 1 iteration termination criteria.

So, while some of the concerns raised regarding curve fitting in a data mining sense are legitimate I don't think that the criticisms of the walk forward adaptation scheme underpinning the OP's approach (there are some plausible ones) make sense.


Thx,
D

p.s. Tradestation and this software are too different for analogies to be drawn between them. It just confuses the discussion.
 
Quote from intradaybill:

James, thanks for the report and for being straighforward. How do you explain this result? I would expect adaptation to result in a better system but you are telling us it fails to do so.
More...

From looking at the parameters james used i suspect that the particularly wide settings used for the long period average led to the performance terrain being too unstable for the swarm to get traction against given the weakly converging adaptation scheme that I touched on in my previous post.

Dakota adaptation is not a panacea (you know this) and has its weaknesses. One of those is that it does not see the parameter landscape in context (not sure if that is the best way to put it). That is, when parameter ranges are wide movement from one adapated parameter set to another can effectively reverse the sense of the system - particularly with simple MA type approaches - but the swarm has no understanding of that.

Also, when paramater ranges are wide, the relative velocity of the bots/agents is higher meaning that the absolute change in value of parameters is higher. So, the bots have higher velocity which usually isn't good. Too little is bad, too much is also bad for reasons which I think are obvious.

if you consider the implication of these two points then you may be able to envisage how the velocity of the bots throws them off (or through) the area(s) of performance they're trying to lock onto. There is also a random element to swarm movement and this is, again, magnified by the size of the parameter space.

When setting parameter ranges wider it may become a defensible approach to inncrease the number of iterations per bar while using a different mechanism to update bot positions. Why? Because if the bots move through parameter space at a consistent though variable rate allowing them to move two or three steps per bar can be justified to permit exploration and to avoid the curse of under-exploration. Another approach might be to allow them to make x iterations every y bars.

I hope I'm making sense and that you find it interesting:)


Thx
D
 
Quote from flyingdutchmen:

it is realy not that hard to understand why this aproach is considered to be optimizing rather than adapting. on this aproach of yours have systems been built allready, a tradestation system that sold for $5k from someone that i could provide his name later to you (i forgot it) comes to mind. his aproach was a socalled "selective optimizing". this system has failed tremendesly in realtime.

now dont get me wrong i am not saying that your performance cant be reached but you are looking back a certain amount of bars to find the best setting that would have gave you the highest performance during this period, this you do all by refering historical data and thus, you optimize (or adapt as you would like to call it) your strategy parameters to those market circumstaces givin in that period.

then the next bar or xx bars later you run this same aproach, look back, find parameters, adjust parameters and run next trade.

i doesnt get more optimized as in your aproach, and what you are doing can be done pretty much in every backtesting software, just not automaticly. that is the only difference imo.

pls explain what is the difference between you running your aproach or me brute force optimizing a tradestation strategy every bar to find the best parameters ?

dont get me wrong, i see how your performace is made out of walk-forward testing and i have no doubt your performace is real and *could be* achieved, but i am critizing your *aproach* and find it anything but adaptive.

i have no idea how one could call your aproach/strategy adaptive


EDIT: the name was dr clayburg
http://tinyurl.com/adaptive-or-not (i used tinyurl here because i feel people like him do not deserve free backlinks to promote their useless stuff)

from his website:
The Universal System is most effective when the system is selectively
optimized when necessary. Selective Optimization is a unique application
of commomly used optimization routines designed to keep the system in
sync with current market conditions.
More...

Thanks for your post. I was considering opening account with TS and buying TTM universal. I have been searching the internet for feedback on this system and found mixed reviews but they were going back to 2008 or 2009.

I am trying to find some automated system with decent result. I have NinjaTrader 7 (not happy with). Have you found a decent system worth using? Thanks in advance.
 
Sam:

It sounds as though you want to buy a system. Most advertised systems aren't super good. Many fail. However, to get a feel for what relatively creditable publicly available systems are working, you can look at the web sites of brokers like http://www.striker.com/En0303_trading_systems_performance_ranking.php?webpageID=8&webmainID=3 It has been years since I've looked at them, but at one time some of the systems could be purchased and run on your machine. In other cases you leased it and it ran on Strikers' site. There was at least one other firm that specialized in running leased/purchased systems for you, but their name escapes me at the moment.

Dr. John Clayburg has been developing and selling systems for decades. I don't consider him to be a snakeoil guy. You can lease some of his systems and then buy them with a credit for some of the lease charges. Systems tend to fail over time. However, the last I looked, Clayburg's Cyclone system was stilling in Future Truth's top 10 ES systems. It was in the 35% annual return range since inception. Not a recommendation as the 35% could be down from 100% two years ago and I no longer subscribe to Futures Truth.

Clayburg develops his systems in Multicharts and ports them to Tradestation. Multicharts runs in a multi-CPU, multi-thread environment. Tradestation, through version 8.x did not. Not sure about 9.x You can get a 30 day free trial to Multicharts. Then it can be leased or purchased. The documentation is very outdated. It is for version 5 and they just released 7. Support is excellent. Depending on your needs, it is an excellent program.

A Google search will get you web addresses for both.

Jack
 
I have also heard some good things about Dr. Clayburg but given that his systems have been cracked and are now available at a fraction of the cost by pirates, I wonder for how long they will maintain their edge.

It is pity. People do a lot of hard work and their systems end up being sold for almost nothing all over the web. This must stop.
 
You must log in or register to reply here.
Back
Top