Factor Investing, Explained

Factor investing is the organizing theory of systematic equity strategies: the claim that the cross-section of expected returns is governed by a handful of factors, characteristics like value, size, momentum, quality, and low volatility, rather than by thousands of idiosyncratic stock stories. If true, alpha is not stock picking; it is disciplined harvesting of factor premia. The subtle and career-defining question is which factors are real and which are statistical mirages.

The linear factor model

Every factor framework starts from a linear return-generating process. An asset's excess return loads on $K$ factors $f_{k,t}$ with sensitivities (betas) $\beta_{i,k}$ :

$r_{i,t} = \alpha_i + \sum_{k=1}^{K} \beta_{i,k}\, f_{k,t} + \varepsilon_{i,t}.$

Under an asset-pricing model, the factors are priced: their expected returns $\lambda_k = \mathbb{E}[f_{k,t}]$ are compensation you cannot diversify away, and the cross-section of average returns satisfies

$\mathbb{E}[r_i] = \sum_{k=1}^{K} \beta_{i,k}\, \lambda_k, \qquad \alpha_i = 0 \text{ for all } i.$

The whole empirical program is: identify factors, estimate loadings by time-series Ordinary Least Squares (OLS), and test whether the pricing errors $\alpha_i$ are jointly zero. A factor "works" if adding it drives the $\alpha$ s of test portfolios to zero.

Risk premia versus anomalies

There are two fundamentally different reasons a factor could earn a positive average return, and they have opposite implications for durability:

Risk premium (rational). The factor pays you because it loses money in bad states of the world, recessions, liquidity crunches, high marginal-utility states. You are compensated for bearing undiversifiable risk, exactly as CAPM compensates market beta. Such a premium is an equilibrium feature and should persist even after everyone knows about it. Value-as-distress-risk is the archetype.
Anomaly (behavioral / frictional). The factor pays you because other investors systematically misprice something, overextrapolation, limited attention, leverage constraints, or institutional mandates. This is closer to alpha, is not "risk" in any utility sense, and should decay as arbitrage capital arrives (Alpha Decay).

The Market Efficiency (The EMH) joint-hypothesis problem means returns data alone usually cannot tell these apart, the same value premium is "risk" to Fama and "mispricing" to Lakonishok. Practically it barely matters why it pays, but it matters enormously whether it will keep paying, and the two stories predict different futures.

How a factor portfolio is built

A tradeable factor is a long-short portfolio that isolates a characteristic. The canonical recipe:

Choose a signal, a firm characteristic $c_i$ (book-to-market, past return, gross profitability).
Rank or z-score it cross-sectionally, often within size or sector buckets, and neutralize unwanted exposures (see Signal Construction).
Go long the top, short the bottom, e.g., long the cheapest 30%, short the most expensive 30%, weighting to be dollar- and beta-neutral.
The factor return is the long-minus-short spread:

$f_t = \frac{1}{|L|}\sum_{i\in L} r_{i,t} - \frac{1}{|S|}\sum_{i\in S} r_{i,t}.$

Because it is long-short, the factor is (to first order) hedged against the market and profits from relative performance, the essence of a Cross-Sectional vs. Time-Series Strategies strategy.

The factor zoo and p-hacking

By 2016 the literature had published hundreds of "significant" factors, Cochrane's "zoo." Harvey, Liu, and Zhu argued most of these are false positives from multiple testing. If you run $M$ independent tests at the conventional $t > 2$ (5% level), the expected number of spurious "discoveries" under a true null is $0.05 M$ , dozens of fake factors from a few hundred tries. Their prescription is a much higher bar: with hundreds of candidates, a Bonferroni-style correction demands roughly

$|t| \gtrsim 3.0 \text{ to } 3.5,$

not 2.0, before you should believe a factor is real. The deeper diagnosis is the same as Overfitting: a signal selected because it had the best backtest is a biased estimate of its true return. Defenses against the zoo:

Out-of-sample and out-of-region tests, a real factor should appear in data it was not discovered in (other countries, other asset classes, pre-discovery samples). "Value and momentum everywhere" is persuasive precisely because it holds across eight markets.
An economic mechanism stated in advance, a risk or behavioral story that predicts the sign before the regression is run.
Deflated Sharpe / higher critical values, explicitly penalizing for the number of trials, as López de Prado advocates.

Worked example: does a factor survive controls?

Suppose a new signal (say, low volatility) earns 5% a year long-short with $t = 2.6$ in a single US sample. Before believing it, you regress its returns on the established factors:

$f^{\text{new}}_t = \alpha + \beta_{\text{MKT}} \text{MKT}_t + \beta_{\text{SMB}}\text{SMB}_t + \beta_{\text{HML}}\text{HML}_t + \beta_{\text{MOM}}\text{MOM}_t + \varepsilon_t.$

If the intercept $\alpha$ is insignificant, the "new factor" is just a repackaging of known ones, it spans nothing. If $\alpha$ stays large and significant and the $t$ -stat clears the multiple-testing hurdle and it replicates out of sample, you have a genuine addition. This spanning regression is the workhorse test of the whole field.

Failure modes

Crowding. Once a factor is famous, capital piles in, valuations compress, and expected returns fall, the low-vol and value trades have both shown this.
Factor timing is hard. Premia are noisy and regime-dependent; trying to time value vs. momentum usually adds turnover and subtracts return.
Definition sensitivity. Small changes in how you build book-to-market or momentum can flip results, a red flag for robustness.
Correlated crashes. Diversified factor books can still crash together in deleveraging events (August 2007 quant quake), because they share the same arbitrageurs.

In interviews

You should be able to write the linear factor model, explain the difference between a risk premium (compensation for bad-state losses, should persist) and an anomaly (mispricing, should decay), and describe how a long-short factor portfolio is constructed. The critique question is now standard: "With hundreds of published factors, how do you know which are real?", answer with multiple testing (raise the $t$ -hurdle to ~3), out-of-sample/out-of-region replication, and a pre-specified economic mechanism. A strong candidate volunteers the spanning regression as the concrete test of whether a "new" factor is incremental. See The Fama-French Factor Models for the models that anchor the whole exercise.

Factor Investing