Choosing the Estimation Universe

A factor model needs a universe of stocks to run its cross-sectional regressions on each period — the estimation universe — and it's tempting to assume this should just be "every stock you might ever trade." It shouldn't be. Include too many tiny, illiquid, or newly listed names and the factor returns and covariance estimates get dominated by noisy, hard-to-trade stocks that barely resemble your actual book; exclude too aggressively and the model is calibrated on a narrow slice that doesn't generalize to the names you actually hold, understating true correlations between your positions and stocks just outside the estimation set.

The standard fix is to build the estimation universe deliberately: rank by market cap and liquidity, keep a broad enough set to estimate stable factor exposures (typically several hundred to a few thousand names, well beyond any single portfolio), but apply floors on price, market cap, and average daily volume to exclude names whose regression weight would be dominated by microstructure noise rather than genuine factor exposure. Some providers use a capped-weighting scheme within the estimation regression itself, so no single giant stock (like the largest mega-cap) can single-handedly set a factor's return that period.

Getting this wrong shows up downstream as a risk model that understates the true volatility of a book concentrated in small caps or one sector, because the estimation universe under-represented exactly those names — a classic case of a risk number being no better than the universe it was estimated on.