Skip to content
All library documents

Distinguishing Systematic Factors from Idiosyncratic Alpha

Article Quant Q&A · Author: ff3
Read the original

Summary

The document explores the distinction between factor research and alpha research. A factor is framed as an exposure that explains a systematic component of returns or risk, while alpha is described as a forecast of idiosyncratic returns without a systematic component. The categories are not absolute: the same feature may be treated as a factor or an alpha depending on the researcher’s horizon, mandate, and context.

One proposed diagnostic is to randomly divide a tradable universe into two groups and construct separate long-short portfolios using the feature in each group. If the feature captures a systematic return component, the two portfolios should tend to move together; a purely idiosyncratic signal would instead produce uncorrelated portfolio returns. The response says familiar factors such as size, value, and momentum display positive correlation under this comparison, while randomized features serve as a contrast. This is a conceptual test, not a complete classification method: many signals combine systematic and idiosyncratic effects, and portfolio construction choices can affect the observed correlation.

Key ideas

  • Factors are intended to capture systematic risk or return components, while alpha targets idiosyncratic returns.
  • The distinction depends partly on an investor’s horizon and mandate.
  • Correlating portfolios formed from separate random subsets can help identify a feature’s systematic component.
  • Many real-world signals lie between clearly systematic factors and purely idiosyncratic alpha.

Tags

EquitiesFactor investingStatistics
Full text
# Difference between factor research and alpha research


# Difference between factor research and alpha research












What exactly is the difference between factor research and alpha research? As I understand it, factor research seeks to explain cross-sectional variation of securities, e.g., does b/p, log(mktcap) at time t explain returns from t to t+N. If that's the case, then the time-series of coefficients of the regression (i.e., the factor returns) is consistently statistically different from zero.

But doesn't that make that factor an "alpha" i.e., a tradable strategy. I often hear that factor model r^2s are high (50s) but alpha r^2s are low (2-5). If that's the case, then what regression are people referring to for the alpha r^2?

## Answer by Chris Taylor (score 6)

https://quant.stackexchange.com/a/83880

There is no clear split between what is a "factor" and what is an "alpha". Something that is an alpha to one researcher might be a factor to a different researcher - depending on time horizon, mandate, firm they work for, level of experience etc. A somewhat tongue-in-cheek definition I have heard is

> An alpha is a factor that my risk manager doesn't know about.

However, we could try to come up with something that is a bit more objective. A factor is supposed to explain a systematic component of risk, whereas an alpha forecasts idiosyncratic returns and does not have any systematic component.

Let's say we have some metric/feature and we want to see if it's more factor-like or more alpha-like. One way to assess this is to split your tradable universe into two, at random, and form long/short portfolios on each subset based on your metric, and calculate the returns for those portfolios.

If the feature loads on some systematic component of risk, then it will be present in both portfolios, and they will be positively correlated. On the other hand, if there is no systematic component then the portfolios will be uncorrelated. You can verify that for well known factors (e.g. size, value, momentum) you will indeed see positive correlations between the two different portfolios, whereas for randomly generated feature values (meant to simulate an alpha with no systematic component) you will get uncorrelated portfolios.

In reality most features will be somewhere between "obviously a factor" and "obviously an alpha" of course.

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.