I am building a gradient descent model based on portfolio optimization. Currently, I have finished the model and am able to run it smoothly without any problem. However, there's one issue that I couldn't fix: the optimal vector components have negative values. As my portfolio does not support short-selling stocks, this optimal solution may not be a convincing solution to my portfolio.

Model Description

My cost function for the portfolio problem is defined as


where $x$ and $\boldsymbol{\Sigma}$ are my minimizer vector and positive definite covariance matrix. $\frac{\gamma}{2}(e^Tx-1)^2+\lambda(e^Tx-1)$ is the augmented lagrangian method for the constraint $e^Tx=1$ and $\rho\sum_{i=1}^n\max(0,-x_i)^2$ is the penalty imposing $x_i>0]$ for all components of $x$.

As can be seen here, my objective is to find the optimum solution for $x$ using the Steepest Descent method. Hence, I start to build my Python code as below:

import numpy as np
from numpy.linalg import norm

def penalty_f(v):
    return  np.sum([np.power(max(0., -xi),2) for xi in v])

def penalty_df(v):
    return -2.*np.array([max(0, -xi) for xi in v])

def aug_lag(lam, v):
    new_lam = lam + gam*(np.sum(v)-1.)
    return new_lam

def lipschitz(v):
    lips = beta_1*np.sqrt(np.trace(cov@cov)) + beta_2*np.sqrt(len(v)) + gam*len(v)
    return min(1,1/lips)

def project_f(v, lam_k):
    func = 0.5*beta_1*v.T@cov@v - b*mean.T@v + gam/2*(np.sum(v)-1.)**2 + lam_k*(np.sum(v)-1.) + beta_2/2*v@v + rho*penalty_f(v)
    return func

def project_df(v, lam_k):
    par_func = beta_1*cov@v - b*mean + gam*np.ones_like(v)*(np.sum(v)-1.) + lam_k*np.ones_like(v) + beta_2*v + rho*penalty_df(v)
    return par_func

def gradient_descent(f, df, ini_v, tolerance, lam, MAX_ITER=10000, output_fname='output.txt'):
    vec = ini_v
    negative = 0
    aug_lam = lam
    with open(output_fname, 'w') as out:
        for i in range(MAX_ITER):
            alpha = lipschitz(vec)
            f_value = f(vec, aug_lam)
            gradient = df(vec, aug_lam)
            direction = np.negative(gradient)
            vec += alpha * direction
            aug_lam = aug_lag(aug_lam, vec)
            print("No. of zeros: ", len(vec)-np.count_nonzero(vec))
            print(norm(gradient,2), "," , f_value)
            msg = f'{f_value},{norm(gradient,2)}, {str(list(vec))}\n'

            # stopping criteria
            if norm(gradient,2) < tolerance:
                print ("The optimum vector for", {df}, " is at ", vec,"at iteration ", i+1)
                for i in vec:
                    if i < 0:
                        negative += 1
                print("No of negative: ", negative)
                print("Vector sum: ",np.sum(vec))
                print("Gradient: ", norm(gradient,2))
            if i == MAX_ITER:
                print ('Higher no. of iterations is needed for', {df})
                print ("Vector: ", vec)
                print("Vector sum: ",np.sum(vec))
                print("Gradient: ", norm(gradient,2))
    return vec

Here is the result I obtained using 33 stocks:

The optimum vector for {<function project_df at 0x000001FA50FF2D30>}  is at  [-0.03442667 -0.01839447  0.03305939 -0.04319195  0.00353058  0.13003235
  0.09405175  0.00022958  0.08010633 -0.02831581 -0.04066535 -0.02875569
  0.04953858 -0.04019486  0.22866438 -0.03474848 -0.00029308  0.09792178
  0.11669689 -0.08569448  0.00405406  0.05544077  0.12273198  0.09615974
 -0.04374827  0.16994986  0.00670517  0.04772238  0.01729612  0.05218545
  0.03484764 -0.03438397 -0.00811171] at iteration  1230
No of negative:  13
Vector sum:  1.0000000001114249
Gradient:  9.961948573544883e-07

When I assess 30 or more stocks, negative components start to appear. Is there anything I can improve or take note of when performing portfolio optimization?

P.S. The data to the mean and covariance is in the link https://www.dropbox.com/scl/fi/rycj948t4bnq5u60m13ow/Covariance.xlsx?dl=0&rlkey=d8u18ntuxk7wjcl1eup8gmxoa

import pandas as pd

df1 = pd.read_excel('Covariance.xlsx', sheet_name=0, header=None)
df2 = pd.read_excel('Covariance.xlsx', sheet_name=1, header=None)
np_cov_1 = df1.values
mean_1 = df2.values.reshape(len(df2))
ini_vec_1 = np.array([1. / (len(df2)) for i in range(len(df1))])
  • 3
    $\begingroup$ Maybe you can simply project after each iteration the negative components onto their limits. You should look into constrained optimization. Python has many suitable algorithms, so that you do not need to reinvent the wheel ! $\endgroup$
    – Laurent90
    Commented Dec 22, 2021 at 9:14
  • $\begingroup$ That's a good idea! But I'm quite a novice to all these optimization algorithms, so if you can show me some directions, that'll be progressive. $\endgroup$ Commented Dec 22, 2021 at 14:50
  • $\begingroup$ Please clarify your specific problem or provide additional details to highlight exactly what you need. As it's currently written, it's hard to tell exactly what you're asking. $\endgroup$
    – Community Bot
    Commented Dec 22, 2021 at 16:55
  • $\begingroup$ What's your value for the penalty parameter $\rho$? The penalty method is only guaranteed to give you a feasible solution (i.e., where the constraint is satisfied) if you let $\rho\to\infty$. $\endgroup$ Commented Dec 22, 2021 at 17:54
  • 2
    $\begingroup$ Here is an excellent paper on nonnegativity constraints. I would recommend the Sequential Coordinate-wise algorithm described, starting at the bottom of page 10. It's quite simple but effective. (BTW, your Python code is very pretty) $\endgroup$
    – greg
    Commented Jan 3, 2022 at 14:23


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.