Using Extreme-Value Portfolios to Analyze Fundamental Factors
Summary
The document explains why researchers may analyze a fundamental factor by forming portfolios from equities at the high and low ends of its values, rather than regressing returns directly on raw factor values across the whole universe. These groups can reveal nonlinear relationships between a factor and returns and may make the resulting model more robust to the choice of factor transformation.
The example is market capitalization: a regression using its raw values can be dominated by a few very large companies. Taking logarithms is one possible adjustment, but the appropriate transformation may be unclear. Comparing portfolios at factor extremes offers another way to study the relationship without relying on a single linear specification. The answer is conceptual rather than empirical; it supplies no tests, performance results, or detailed portfolio-construction rules, and it does not claim that extreme portfolios eliminate the need to consider factor scaling or other modeling choices.
Key ideas
- Portfolios formed from high and low factor values can help reveal nonlinear return relationships.
- Extreme-value groupings may make factor analysis more robust to transformation choices.
- Raw market capitalization can let a few very large companies dominate an ordinary least squares regression.
- Logarithmic scaling is one possible treatment, but the best transformation may not be obvious.
Tags
Full text
# Fundamental factor analysis using portfolio construction # Fundamental factor analysis using portfolio construction I am a new aspiring quant who is trying to build a fundamental factor algorithm to rank stocks for a basic long/short strategy, so sorry for the likely very basic question. Nevertheless... Why do you need to regress fundamental factors you would like to test against portfolio returns constructed of equities of the highest and lowest extremes of those values in your trading universe. Why not just regress the factors against the returns of your whole trading universe? Is this just because it is too computationally expensive or does the construction of those groups of equities serve another function? Cheers ## Answer by Tim Wilding (score 1) https://quant.stackexchange.com/a/40865 If I understand you correctly, you don’t need to build the groupings, but the construction of the groups of equities allows you to account for any potentially non-linear effects in the response of equities to your factor. It can also help to make your model more robust. For example, people often talk about a size factor, but using raw market capitalisation would give you a factor that is dominated by one or two names (Apple,Google, etc. in the US market) if you use OLS regression. It is not clear that there is a linear response to a market cap factor so people often use log(market cap), but that choice may not be so obvious. That leaves you with a choice of how to transform your factor. Given that, it may be better to go via these baskets of extreme securities to model the factor.
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.