Conditional value-at-risk bug

[robertmartin8]

Currently, the CVaR optimisation using NoisyOpt is a little buggy. Because the weights aren't normalised by default, we must post-process them. However, this post-processing also means that the final weights don't respect the initial bounds. I'd appreciate any suggestions for a fix.

[robertmartin8]

I have received a lot of feedback, and it is clear to me that my current implementation of CVaR needs to be greatly improved or removed. Right now, we basically have a negative_cvar_objective that computes the CVaR from 10,000 runs – this objective is then naively optimised via noisyopt.

I am aware that CVaR optimisation can be formulated as a linear programming problem, so I would like to rewrite the code to do that, and remove the noisyopt optimisation, which has proven to be quite unstable.

http://artax.karlin.mff.cuni.cz/~branm1am/download/VAO_HW1.pdf
https://www.ise.ufl.edu/uryasev/files/2011/11/kro_CVaR.pdf
http://www.diva-portal.org/smash/get/diva2:479147/FULLTEXT01.pdf

[farbodbahari]

Hi Robert, Are you still considering implementing the CVaR? I have used nonconvex_objective method from EfficientFrontier and then minimized CVaR (Negative CVaR) function via SLSQP solver. It appears to me it is properly working. I would be happy to send you my code and see it yourself if you'd like to.

[robertmartin8]

Hi @farbodbahari,

I'd love to take a look if you don't mind sharing! That said, CVaR can be implemented as a convex problem, so realistically I won't be adding CVaR to the library until I've figured that out (or someone else PRs it).