PyGAD Question #102
Unanswered
Gus-Magalhaes
asked this question in
Q&A
PyGAD Question
#102
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Hello all
My name is Gus. I belong to a group of amateur trader who are testing trading strategies in the futures market. For this, we are using the PyGAD library in order to obtain the best parameters for each strategy. Thank you so much for making such code available, it's been very helpful! I believe we are in trouble and would appreciate it if you could give us some thoughts as to what we might want to try.
One problem is that we are not able to make fair comparisons between strategies, given that the fitness results are very different each time we run the same combination of parameters of the genetic algorithm (GA). As a result, we are unsure as to whether a run down is:
Before changing a GA parameter, the same combination is executed several times. In this we observe if there is an improvement in the fitness average (Log of accumulated profitability). However, I fear that the variability of final results is too great and consequently, we would need many runs to have an average fitness representative of the tested parameter combination. This would lead to a lot of processing time and working hours.
Another problem is how to change the parameters: We aim to calibrate the GA so that the final results for the same combination of parameters have little variability between them. This objective would help us to have better decision quality assurance when comparing two different combinations of parameters (varying one by one). Another objective is to calibrate the GA so that the final results have a high average fitness (high profitability).
These mentioned objectives may or may not be conflicting. To try to solve these problems, we are trying to put a second GA to adjust the first GA. This can even resolve the effect of bias when testing parameters one by one, as this would not consider their interaction effect. Hence, some doubts arise: what parameters would this second GA have? What would the fitness be? How do we reduce the variability present in the final results of the GA for the same combination (without having to fix the random seed)?(... Is this even possible?) Our GA 1 fitness go is not in the form 1.0/(abs(output - desired_output), does that matter?
Thank you very much!
Best,
Gus
Beta Was this translation helpful? Give feedback.
All reactions