Regularizing Global Minimum-Variance Portfolio Weights
Summary
The document explains how regularization can stabilize global minimum-variance portfolio weights. In the unconstrained case, the optimization objective can penalize large weights using an L2 norm, as in ridge regression, or an L1 norm, as in lasso. This discourages extreme offsetting long and short positions. When short selling is prohibited, the weight constraints already limit some extremes; comparison with an equal-weight portfolio can provide a further benchmark against sparse allocations.
A tuning parameter controls the strength of the penalty. The suggested approach is to compare portfolio variance across regularizers and choose the parameter using validation or cross-validation. The discussion is conceptual and does not report results on the cited daily industry-portfolio dataset, prescribe a validation split, or establish which penalty performs best. Those choices require empirical testing and should account for the time-series structure of return data.
Key ideas
- Global minimum-variance optimization chooses asset weights subject to a portfolio-weight constraint.
- L2 and L1 penalties discourage extreme weights in different ways.
- Long-only constraints already restrict the solution, and equal weighting can serve as a comparison benchmark.
- A regularization hyperparameter can be selected using validation, then compared by portfolio variance.
Tags
Full text
# Compare portfolio variance using different regularizers
# Compare portfolio variance using different regularizers
I'm given a question like below. Using the 48_Industry_Portfolios_daily dataset: characterize/describe the dataset and focus on the global minimum variance portfolio. Compare the portfolio variance using different regularizers and use validation methods to find the optimal parameters.
What I'm not clear is to compare the portfolio variance with different regularizers and to use validation methods.
I was using Python to find the efficient frontier. What I need to know is, are there any useful python materials where I can compare portfolio variance using different regularizes. I was not able to find useful resources
P.S : Efficient frontier doesn't look good
## Answer by Attack68 (score 3)
https://quant.stackexchange.com/a/44325
When you solve for a minimum variance portfolio you acquire some values, $\mathbf{\beta}$ corresponding to the weights of your assets, usually such that $\sum \mathbf{\beta} = 1$.
Regularization means you try to limit these values such that your objective function also includes the norm of $\mathbf{\beta}$ (Ridge regression - L2-norm) or the sum of absolute values of $\mathbf{\beta}$ (Lasso - L1-norm). If you allow short selling this means you will try to avoid the case where one asset might have a weight of -1000 and another +1000.
If you don't allow short selling the assets already contain some sort of regularisation but you might further control it by comparing it to the case, for example, where each asset is weighted equally, to avoid many assets having zero weight.
Usually you will control the amount of regularisation with a hyper parameter assigning how much weight to the ridge (or lasso) component you want to prioritise. If you review the documentation for Python's SKlearn Library you will find a lot of documentation about regularisation, as well as cross-validation which is what you will need to test your choicesShown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.