Most adaptive trading systems seem static. You specify the parameters and they spit out an optimized strategy with the highest sharpe ratio (and you're pretty much stuck with it). But the lingering question that remains is whether it can adapt to the changing market conditions as they unfold. Hence my question: is there any that can change the parameters on the fly? For example, I throw in as many as 20 different strategies and the system is able to not only choose the most suitable one for the current market condition but is able to dynamically optimize varying parameters.
Any machine learning trading system with adaptive parameters?
- Thread starter schizo
- Start date
Any online learning system should be capable of doing this e.g. stochastic gradient descent. What you describe in terms of combining the predictions from the individual strategies is referred to as boosting in the machine learning field. You might find this interesting to read into.
You could also look into Kalman filters which will let you adaptively track system parameters over time.
You could also look into Kalman filters which will let you adaptively track system parameters over time.
Yes, it is called Artificial Intelligence. Googled it. But I warn you it is heavy stuff.
Basically you should feed the machine with information and the machine will calculate results of all possible combinations and is able to learn from these calculations and improve the result. We speak about billions and billions of calculations on systems that run 24/24.
I once bought such a package in an euphoric mood. Had to find a PHD in math's to try to define what to feed, how to feed, how to define variable settings......
It was unfortunately not a black box that would finally give you a ready to use system. After a lot of trials we gave up and threw everything in the garbage container.
If AI would be easy or give any result at all, all the financial strong companies would use it. And this is not the case, so....
Last edited:
FYI: Michael Himmel seems to be in this space, his url: http://www.essex.us.com/about
I have not seen any publications on any details, however.
I have not seen any publications on any details, however.
Running 20 separate backtests prior to updating the system is not doing it on the fly. Very diffficult to do with 5-minute data, you may be constrained to eod trading.
Last edited:
If say a given strategy depends on the size of an indicator (e.g., ADX), it might make more sense to choose the parameter(s) that maximize the average ADX value over the N most recent bars rather than go through the whole backtest routine.
Last edited:
That seems a sensible approach.
Thanks for all the replies.
Is anyone familiar with BioComp Dakota 3? This comes close to what I've described above. Here's a brief description from their website:
Why is Dakota Different?
BioComp Dakota uses "Swarm Technology"(tm) to adapt the parameters of programmed trading systems to shifting market conditions in an attempt to create and maintain profitable market timing signals, that is, when to buy, sell or exit from the markets in a profitable manner. Dakota uses Swarm adaptation technology, not optimization. You don't want Swarm Optimization which "jumps" from one set of trading system parameters to another, but adaptation, where the swarm tracks performance smoothly. Systems with Swarm Adaptation algorithms are rare as most Swarm Technologies are focused on ill-fated optimization.
Is anyone familiar with BioComp Dakota 3? This comes close to what I've described above. Here's a brief description from their website:
Why is Dakota Different?
BioComp Dakota uses "Swarm Technology"(tm) to adapt the parameters of programmed trading systems to shifting market conditions in an attempt to create and maintain profitable market timing signals, that is, when to buy, sell or exit from the markets in a profitable manner. Dakota uses Swarm adaptation technology, not optimization. You don't want Swarm Optimization which "jumps" from one set of trading system parameters to another, but adaptation, where the swarm tracks performance smoothly. Systems with Swarm Adaptation algorithms are rare as most Swarm Technologies are focused on ill-fated optimization.
Interesting thread. Make sure you're aware of your bias/variance trade off. The potential to overfit and introduce too much bias is incredibly high here. Otherwise, yes. No reason not to use an online algorithm to modify your model parameters as long as you use sensible validations of parameters and understand bias vs. variance. If you can arrive at a mathematic derivation of the optimization problem (look up maximum likelihood estimator) this can sometimes tell you something meaningful about how your strategy performs in different situations because you can assess what values parameters take on at critical points. This is often not possible in practice...the analytical approach, that is.