Factor Research

The idea

From stocks to characteristics

Factor investing begins with an empirical observation that will not go away: large groups of stocks sharing a characteristic — cheapness, recent momentum, small size, quality balance sheets, low volatility — have historically earned returns that a market beta alone cannot explain. The modern framework descends from the CAPM through Ross's arbitrage pricing theory to Fama and French: expected returns are compensation for exposure to a small number of systematic risk dimensions,

where \(\beta_{ik}\) is asset \(i\)'s loading on factor \(k\) and \(\lambda_k\) is that factor's premium. The canonical six:

Market (MKT)

The equity risk premium itself — exposure to aggregate market movements. Everything else is defined net of it.

Size (SMB)

Small minus big. The premium is historically real but episodic, and much of it concentrates in the smallest, least liquid names.

Value (HML)

High book-to-market minus low. The oldest documented premium and the most publicly tortured — a decade-plus drawdown after 2007 became the field's favorite argument about whether premia die or hibernate.

Momentum (UMD)

Recent winners minus recent losers. Strong average returns punctuated by violent crashes when the market reverses — a premium that looks like selling insurance against regime change.

Quality (QMJ)

Profitable, stable, conservatively financed firms minus their opposites. Robust across specifications, and suspiciously close to "things Warren Buffett already told you."

Low volatility

The anomaly that inverts theory: lower-risk stocks have earned higher risk-adjusted returns, plausibly because leverage-constrained investors overpay for lottery-like volatility.

Methodology

Cross-sectional regression

The standard estimation machinery is the Fama-MacBeth two-pass procedure. First, time-series regressions estimate each asset's loadings:

Second, at each date \(t\), a cross-sectional regression of returns on those loadings estimates the premia:

The time series of \(\gamma_{k,t}\) gives both the premium estimate and — because the cross-sectional estimates are (roughly) independent across dates — its standard error. The procedure's honesty comes from that second pass: a factor must be paid in the cross section, date after date, not merely correlate with returns in aggregate. Its dishonesty risks are equally well catalogued: errors-in-variables from estimated betas, premia that vanish outside the sample, and a research industry running thousands of implicit hypothesis tests on one shared history. Harvey, Liu, and Zhu's t-statistic threshold of 3.0 exists because the factor zoo is largely a multiple-comparisons problem wearing a lab coat.