Factor exposure

Before a stock's return can be called a judgment on the company, the part that any stock of its kind would have earned has to be removed. The factor regression does that removal. It explains daily excess returns with the market premium and five standard style factors, and what is left over is the part specific to the company.

What it measures#

Daily returns in excess of the risk-free rate are regressed on the Fama–French five factors (market, size, value, profitability, investment) plus momentum. The coefficients are the stock's exposures; the intercept, annualized, is the return not explained by them; the R-squared is the share of variance the factors explain, and one minus it is the idiosyncratic share. Standard errors are Newey–West, so autocorrelation and changing volatility do not flatter the confidence intervals. Rolling windows show whether the exposures have been stable, a chronological split shows whether the fit survives out of sample, and the seven tests in each dossier are corrected for multiple comparisons.

Assumptions#

The factor files come from the Kenneth R. French Data Library and end one to two months before the price file; the regression uses the overlap and publishes both end dates. Returns are split-adjusted price returns unless a dividend table is pinned. Linearity and constant exposures within a window are assumed, which is exactly what the rolling and out-of-sample checks are there to test.

When it misleads#

A newly listed company has a short sample, and a short sample produces wide intervals; the artifact flags it and the memo does not read a precise beta into it. An intercept that looks large is rarely significant once the sample and the multiple-testing correction are respected, and the lab publishes the corrected value rather than the raw one. Factor exposures also describe the past sample, not a law; a company that changes its business changes its loadings.

How the lab uses it#

The factor section tells the reader how much of the stock's story is the AI trade at large and how much is this company, and it sizes the idiosyncratic risk the thesis is actually taking. The demo below fits the same estimator to a synthetic series whose true exposures are known, so the reader can see what a recovered coefficient and its interval look like when the answer is certain.