# [R-SIG-Finance] Quadratic programming solve.QP's Lagrangians

Ton Jean-Claude jean-claude.ton at epfl.ch
Wed May 27 14:41:52 CEST 2015

Hi I am currently playing a little with Markowitz's portfolio optimization problem and from my computations I get that if we start from the following problem :

$$\underset{w}{\arg \max} \quad w'r - \frac{\gamma}{2} w' Q w \quad \text{s.t} \sum_i w_i=1, \quad w_i \geq 0\quad$$

Then $\gamma$ should be of the form :
$$\gamma = \frac{w'(r-\beta)-\alpha}{w' \Sigma w},$$

where $\alpha$ is the Lagrangian multiplier due to the sum of weights being equal to one and $\beta$ the Lagrangian multipliers due to the no short selling condition.
Using solve.QP from the package quadprog I have coded the mean-variance formulation. And the function ComputeRho which you may find attached below computes the $\gamma$ (Sorry for the bad notations)

My problem is the following
depending on the r.a (risk aversion that is $\gamma$ in my case), ComputeRho gets me the correct answer or not.

rm(list=ls())
###################################### begin : packages      ####################################
require(tseries)
require(quantmod)
require(MASS)
require(PerformanceAnalytics)
require(corpcor)
require(quantmod)
require(parallel)
require(ggplot2)
#require(mvtnorm)
###################################### end : packages  ######################################
LogRet <- function(x,lag=5){
#require(quantmod)
#return(na.omit(apply(x,2,function(x) Delt(x,k=lag))) )
require(zoo)
return(na.omit(diff(log(x),lag=lag)))
}
##########################################################################
# Input :        sol = solution from solve.qp
#                lr = logreturns
# Output :      rho
ComputeRho<-function(sol,lr){
lr<-as.matrix(lr)
alpha<-sol$lagrange_mults[1] # Lagrangian due to normalizing constraint beta<-sol$lagrange_mults[2:length(sol$lagrange_mults)] # Lagrangians due to no short sell constraint weights<-sol$weights
cov.mat<-cov(lr)
# unconstrained form
uncons <-(t(weights)%*%colMeans(lr))/(t(weights)%*%cov.mat%*%weights)
# due to Lagrangian multipliers
cons1 <- (t(weights)%*%beta)/(t(weights)%*%cov.mat%*%weights)
cons2 <- alpha/(t(weights)%*%cov.mat%*%weights)
cons<-cons1+cons2

rho<-uncons+cons # lagrangians change sign...
rho
}

############################################################
# Input :       data = usually logreturns
#               r.aversion = risk aversion coefficient usually between 1 to 10
# Output :      er = expected return
#               sd = standard deviation
#               weights = weights
#               Lagrangians = Lagrange multipliers

MeanVariance<-function(data,r.aversion=10,method="MeanVariance")
{
# compute probability beliefs mean and risk ( here sample covariance)
er<-colMeans(data)
cov.mat<-cov(data)
# get length of data
n_asset <- length(er)

Dmat <- r.aversion*as.matrix(cov.mat)
dvec <- er
meq <-  1

bvec <- c(1, rep(0,n_asset))
Amat <- matrix( c(rep(1,n_asset), diag(nrow=n_asset)), nrow=n_asset)

sol<-solve.QP(Dmat, dvec,Amat, bvec=bvec, meq=meq)

# fetch returns from solve.QP
weights<-sol$solution exp.ret <- t(er)%*%weights std.dev <- sqrt(weights %*% cov.mat %*% weights) # return results ret <- list(er = as.vector(exp.ret), sd = as.vector(std.dev), weights = weights, lagrange_mults=sol$Lagrangian,
weights2=sol\$unconstrained.solution)

}
# risk aversion
r.a<- 17 # for values below 18 r.a is different from rho
lr<-LogRet(as.matrix(data))

sol<-MeanVariance(lr,r.aversion = r.a)
rho<-ComputeRho(sol,lr)
rho

r.a2<- 18 # for values below 18 r.a is different from rho
sol2<-MeanVariance(lr,r.aversion = r.a2)
rho2<-ComputeRho(sol2,lr)
rho2

I was able to cern the problem to the sign of the Lagranians (here cons1 and cons2 in ComputeRho ) which switch signs depending on the r.a. I choose.

Does someone knows how the Lagrangians are computed explicitly ? Since It seems that this is done in FORTRAN.

Best,
Jc

here is the data set
https://www.dropbox.com/s/ynlnmdgfm9w4rq0/data.csv?dl=0

[[alternative HTML version deleted]]