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.
this socalled "adaptation"engine, which, in my eyes is still curve fitting on the fly, got my attention and this interest has gone as far as building my own engine in tradestation. probable not as sophisticated as the one we are discussing here, but as far as i can see it seems to do pretty much the same job. i have one strategy which is very dependent on changes in volatility, it has one parameter, instead of trying to forecast volatility i will try to let this litle engine find out what are the best settings for n based on a swarm of invisible indikators that run on the background and store their data into a bunch of arrays over x amount of theoretical trades. its not realy something i usually would do but it seemed like a great oppertunity to enhance my EL knowledge. due to the multiple calcuations on the background the script is pretty slow but i didnt expect otherwhise. more next week, this thingy made flyingdutchmen curious
I've been thinking about it, but I can't really add anything other than my approach didn't 'work' in this case. The exercise prompted the following thoughts though. Is it possible to adapt the SMA Crossover system rapidly enough to produce a good result? In an effort to discover if it is possible, create a series that represents close to the ideal path for the adapted parameters. Adapting only the longer SMA Period would do. The path would be along the lines of minimize the degree of change in the longer SMA and maximize the performance. This would be done using a good variety of markets and periods. Now, analyze / graph information that is not being used to identify any performance variables that would do the trick. In Dakota the performance information is fed to the swarm of bots from the Equity Engine (performance engine). Where the bots move to in the parameter space depends heavily on this information. In previous posts I have mentioned the idea of bots moving to areas in the parameter / performance space where performance has been historically rising, rather than where it has been best. So the swarm of bots might move to an area that has been historically unprofitable, however, performance delta is increasing. Maybe an approach along these lines would result in a performance engine that is able to anticipate regions in the parameter space that will be profitable to a better degree than what I am doing now. FlyingDutchman, great to hear that you are giving it a go. Regarding the adaptation versus walk-forward curve fitting. I think walk-forward curve/function fitting of the type that we are talking-about is a type of adaptation. I can't think of any type of adaptation that doesn't involve some type of curve fitting. Categorizing the different ways that we can build adaptive systems is an exercise that I have been planning to start on. Best Regards, James
just curious, how man times did it occure that you where dissapointed in the results after adding dakota to an strategy and would you think this approach would suit better to one-parameter strategies than multiple-parameter strategies ? can you find any correlation to what kind/type of strategy would do better out of sample after adding dakota to it ?
This has been a very informative thread...nice Simple comment....it seems that while there may be advantages in 'adaptive systems' the end of day outcome appears that simple one parameter systems fare equally as well... One thing that i would like to see for all the provided results would be the maximum consecutive losers/winners and drawdown... Any possibility this could addressed ;especially with the 5>30 30<5 system results? NiN
The SMA Crossover 5/30 system requires rapid adaptation of parameter values. Possibly a total inversion of the model is required aperiodically. The systems that I typically build use a lookback period of 4 years and the parameter values tend to move smoothly around the parameter space. The challenge for me now is to see if I can build a performance engine that will enable the bots to anticipate profitable regions in the parameter space rather than simply move to areas that have been historically profitable. Initially, I will give my performance delta idea a run and see if it has legs. This won't take long and I will post the results. Coincidentally, I have been re-writing my reporting engine over the last week or so. The quality of the reports that I produce will be greatly improved soon (the current reports aren't too flash). Best Regards, James
I haven't tended to build systems based on published strategies. I have one system that uses David Varadi's DVAM indicator and I have built a system similar to the 11 period SMA Diff EMA indicator. For the adaptive version of the SMA Diff EMA I set the minimum periods to 3 and max to about 20. A year or so ago I would used wide ranges for all parameters. A rule of thumb that I now use is to only adapt what needs to be adapted for a given system. The best case scenario would be a system that handles all market regimes with no adaptation. The second best case would be a system that handles all market regimes using one adapted parameter. For the approach that I am currently taking (4 year lookback period), systems that benefit from relatively subtle bar by bar adjustments in the adapted parameter values perform better whether running a trading simulation or using the system for trading post construction. Best Regards, James
I was challenged to demonstrate the efficacy of my own adaptive moving average so here are my test results.
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.