You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
like we did in the R package, we should show examples where general solvers (from e.g., scipy.optimize) are not able to or are too slow to solve the non-convex risk parity formulation.
The text was updated successfully, but these errors were encountered:
I'm a master program CS student and I was searching in jax dependency graph for projects using it to learn it's usage.
I came across this repository and after inspecting it I would like to learn more and try contributing.
Is this issue free to take?
As I understand the goal here is to show performance comparison between SCRIP and scipy.optimize solvers, similarly to the example in riskParityPortfolio documentation for risk concentration only example. Is that correct?
Following the derivation from presentation I used SLSQP solver from scipy.optimize because it's the only one that supports both equality and inequality constraints.
I made a PoC notebook, where for trick case used in tests I ran scipy solver.
Although for initial points that are close to the solution it does converge, it fails for others. I used initial guess from newton_nesterov.cc from riskParityPortfolio project:
like we did in the R package, we should show examples where general solvers (from e.g.,
scipy.optimize
) are not able to or are too slow to solve the non-convex risk parity formulation.The text was updated successfully, but these errors were encountered: