sum(abs(weights[i])) == 1

[dmicsa]

I don't find a way to add a constrain that forces my total absolute weight to be 1. I've tried all range of lambda like this one, for example:

ef.add_constraint(lambda x : sum(abs(x)) == 1)

I want the total exposure to be 1 and to support negative weights too.

[robertmartin8]

Hi @dmicsa,

Thanks for reaching out! Is this different from taking the weights and rescaling them as required?

ef = EfficientFrontier(mu, S, weight_bounds=(None, None))
ef.max_sharpe()

rescaled = ef.weights / np.abs(ef.weights).sum()
np.abs(rescaled).sum()  # = 1.000000

I've had trouble in the past adding explicit L1 norm constraints (not convex).

Best, Robert

[dmicsa]

It came to my mind to rescale them after but after what I saw something hugely hedged like this: Weights: {'tmf': 0.21, 'vnq': 0.07, 'qqq': 0.69, 'mdy': -0.68, 'vwo': -0.1, 'epp': -0.11, 'ilf': -0.24..} as the main holdings, I dont quite like 'qqq': 0.69, 'mdy': -0.68 part, I said better to ask. :)

btw your PyPortfolioOpt is an amazing library I tyvm for it!

Too bad is nearly impossible to install in Windows, I had to install Windows Subsystem for Linux to have it running and is darn clumsy to use it as a Windows user.

Your solution doesnt work quite so simple: ----> 4 rescaled = weights / np.abs(weights).sum() 5 np.abs(rescaled).sum() # = 1.000000 6 print('Weights:', rescaled)

TypeError: bad operand type for abs(): 'dict'

I had to write some ugly plumbing but is working.

def rescaleWeights(ws): s = 0 for key in ws: s += abs(ws[key])

print('s = ', s)

for key in ws:        
    ws[key] /= s
    ws[key] = int(ws[key] * 100) / 100    

[robertmartin8]

Hi @dmicsa,

It came to my mind to rescale them after but after what I saw something hugely hedged like this

Unfortunately, this is an issue with mean-variance optimisation as a whole – it can lead to very unrealistic weights unless you put many constraints on the problem. If you want more stable portfolios, you may consider ef.min_volatility() with an L2 reg objective:

ef = EfficientFrontier(mu, S, weight_bounds=(None, None))
# experiment with different values of `gamma`
ef.add_objective(objective_functions.L2_reg, gamma=1)
ef.min_volatility()

# rescale
rescaled = ef.weights / np.abs(ef.weights).sum()
# to convert to a nice dict
weights = dict(zip(ef.tickers, rescaled))

Too bad is nearly impossible to install in Windows, I had to install Windows Subsystem for Linux to have it running and is darn clumsy to use it as a Windows user.

I wasn't aware about that! Do you have any more information about this? For example, does a simple pip install PyPortfolioOpt or poetry add PyPortfolioOpt not work?

Your solution doesnt work quite so simple:

In the snippet above I am using ef.weights, which is a numpy array, instead of weights, which is a dict, so my rescaling snippet should work fine.

Best, Robert

[dmicsa]

Thank you Robert for help.

I give it a try on some universes with: ef.add_objective(objective_functions.L2_reg, gamma=1)

and I always receive: OptimizationError: Please check your objectives/constraints or use a different solver.

Maybe I'm not that skilled to fix this.

I tried to install it on Windows again but I waisted over 10 hours trying around 1..2 months ago. I tried today too and again problems. Maybe I will investigate it when I will have some more time.

I received this error today:

  Running setup.py clean for scs
Failed to build scs
Installing collected packages: scs
    Running setup.py install for scs ... error

On Linux subsystem installs clean but it is annoying to don't have it natively on Windows.

I use it because many other user had the same problems installing it on Windows and they suggest this.

[dmicsa]

I managed to install it under windows it helped a lot:

conda install -c cvxgrp cvxpy

from: https://anaconda.org/cvxgrp/cvxpy

Now I have already on my system on Win10: pyportfolioopt-1.0.1-py3.7

So if I run: pip install PyPortfolioOpt will do nothing or will not update to the latest version. Is it possible to do it on windows like that and have the latest version?

Dan

[dmicsa]

TYVM for your time. I reinstall Conda and everything from scratch and managed to install it in windows after 10 hours of wrestling. :).

[giusscr]

Hi Robert, it has been a month since i'm stuck with the problem of setting the sum of absolute weights= 1, I think that I found a raw solution and I would like to give my contribution, I'm a beginner user of python and GitHub as I set this account up just for this matter.

Please let me know how can I contact you so we can discuss it together.

[88d52bdba0366127fffca9dfa93895]

Hi @giusscr, you could post your suggestion here so anyone could discuss and review your work.

[giusscr]

Sure, I made a test that I think that it is harder to explain than seeing thats why I attached a colab file.

For a project I'm doing with university I need to create a long/short portfolio without leverege so that the abs sum of the weights is equal to 1.

After almost a month of trial and error ( I'm new to python) I tried to test to see what would happen if I made a simple optimization ( weights bounds [0,1]) by using as input:

1) the expected return all positive (even for the stocks i want to short)

2) a cov_matrix derived from a series of stock returns where i change the signs of the stocks i want to short( I call it in the file "stock_returns_inverted".

Obviously i would get all the weights positive, so i changed the sign of the weights that i wanted to short and then I used the base optimizer to see if i would get the same portfolio performance and volatility using as input this new set of weights, the original expected returns and original cov_matrix.

And indeed it works. I also made a comparison of performace with the rescaled solution and i get a better sharpe ratio.

In the file I used this set of stoks ["ADBE", "AMZN", "ENPH", "FTNT", "MCD", "PFE"] with this set of expectations er= [-0.15, 0.08, 0.30, 0.14, 0.12, -0.20].

https://colab.research.google.com/github/giusscr/test-long-short-portfolio/blob/main/test_to_send_github.ipynb

[giusscr]

@88d52bdba0366127fffca9dfa93895 @robertmartin8 what do you think? Is it financially sound or did I just got lucky with the numbers ?