Using Lasso and PCA to Select Market Return Predictors

Summary

The document asks how to identify a useful subset of roughly one hundred market related variables for explaining or predicting global stock returns. It highlights two data settings: fewer observations than predictors and at least as many observations as predictors. The questioner favors Lasso for its simplicity and notes that principal components can be difficult to interpret.

The answer points to research using Lasso for asset pricing and high dimensional model selection, alongside studies applying PCA. These references show that both approaches have been studied in financial prediction problems, but the document does not compare their performance or prescribe a selection procedure. It provides no empirical results, implementation details, or guidance on handling correlated predictors, choosing tuning parameters, or validating selected variables. The cited literature is an entry point for further reading, not evidence that either method will isolate independent causal drivers of global equity returns.

Key ideas

• Lasso is presented as a candidate for selecting predictors in high dimensional asset pricing problems.
• Principal component analysis is another cited approach, though its components may be harder to interpret.
• The question distinguishes settings where the sample size is below or above the number of predictors.
• The answer names research references but does not compare methods or establish that selected variables are causal drivers.

Tags

# Subset selection to identify independent variables that impact the market?


# Subset selection to identify independent variables that impact the market?












Given a lot of market-related features (~100 independent variables such as emerging market, developed market, s&p 500, tech sector returns, etc), I need to select a subset of them that are ideally independent and are the major drivers of the global stock market return during time t=t1 to t=t2.

Specifically, the model has to identify important/non-important variables when: 1) the number of independent variables (p) is large (~100) 2) the number of sample size (n) < the number of independent variables (p) and when n >= p

Are Lasso and PCA good ways to accomplish this? I guess Lasso is a simple, easy method. I think that the problem with PCA is that the interpretation of the result is not going to be easy...

Are there academic literature that deals with this problem (selecting a subset of large independent variables to predict the global stock market return)

## Answer by phdstudent (score 2, accepted)

https://quant.stackexchange.com/a/40303

The literature on Lasso for asset pricing is quite recent and there are few references out there yet. The main ones are:

- Freyberger, Neuhierl, Weber - Dissecting Characteristics Nonparametrically - this uses Lasso.

- Huang and Shi (2016) - also lasso.

- Horowitz (2016) - gives a overview of model selection in high dimensional models

Also several papers on PCA:

- Giglio and Xiu (2016)

- Kelly, Pruitt, and Su (2017)

Shown 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.