Menu

#7 Hyper parameter search

Version 1.1
open
nobody
2022-11-12
2022-01-22
Anonymous
No

Originally created by: jbaron

Most trading strategies are non-differentiable, but they have a lot of parameters that need to be tuned. Most simple hyper parameters searches take too much time, so research is required for viable alternatives.

Discussion

  • Anonymous

    Anonymous - 2022-03-05

    Originally posted by: mpaoletta

    Hi, we wrote a simple grid search that runs backtests in parallel, using a cached feed in order to accelerate the search. It's not optimized in any way regarding following a path of better results (like running some tree optimized search - don't know if it makes any sense...)

     
  • Anonymous

    Anonymous - 2022-03-06

    Originally posted by: jbaron

    Might be a good starting point. FYI, my ideal end goal would look something like this (not all for shorter term):

    • be able to define both discrete and continuous search spaces per parameter.
    • be able to define parameter constraints (like param1 has to be bigger than param2)
    • support for both basic (grid, random) and advanced (bayesian, ...) search optimisers
    • run across multiple machines to speed things up

    But right now I personally just write hard-coded nested loops, so any start will be an improvement on that ;)

     

Log in to post a comment.