Factor Investing

Scope. This article explains what a factor is, what factor investing claims, and why the claims are unusually hard to establish. It recommends no factor, names no product, supplies no screen or weighting, and does not rank factors against each other. The statistical hazards are set out at greater length in Pillar 25's article on quantitative analysis, which this one cites rather than repeats.

Size, valuation ratios, past returns, profitability, investment intensity and volatility are the usual examples.

Factor investing is the decision to take deliberate exposure to one or more of those characteristics rather than to individual companies. The unit of analysis moves from the business to the attribute, which is a genuine conceptual shift rather than a repackaging.

Two things called by one name

Factors as description and factors as prediction are different activities, and conflating them is the source of most confusion in the subject.

As description, factors are uncontroversial and useful. Any portfolio has exposures to measurable characteristics whether or not anyone chose them — a collection of holdings assembled entirely on other grounds will still tilt toward or away from smaller companies, cheaper ones, more profitable ones. Measuring that is diagnosis, and it tells an investor what they own in terms other than a list of names. Nothing in the criticism below touches this use.

As prediction, factors are a claim that a characteristic carries a persistent return premium. That is an empirical assertion about the future, and it is where the difficulties start.

Why establishing a premium is harder than it looks

A factor premium is a small number surrounded by a large one. Premia are typically a few percentage points a year; the volatility around them is typically ten to twenty percentage points. Distinguishing the first from zero, given the second, requires a great deal of data — and the amount required can be computed exactly.

Worked example — the evidence required arrives more slowly than the careers that need it. A 3% premium with 10% volatility needs 44 years of data to reach a t-statistic of 2.0, and 100 years to reach 3.0. That second number is the one that matters, because the study cataloguing the published factors argued that a hurdle above 3.0 is the appropriate bar precisely because so many factors have been searched — as Pillar 25 sets out, 316 factors had been catalogued and roughly 426 would clear a conventional bar by chance alone from a search of that size. So the field's own recommended standard requires, for a typical premium, about a century of data that mostly does not exist. This is not an argument that factors are fictitious. It is the reason honest disagreement persists: the samples available are too short to settle the question by the standard the field itself proposes. (The years-to-t figure is (t × volatility ÷ premium)², the same construction as the alpha-detection years in Pillar 22.)

The definitional problem, which is separate and just as awkward

A factor is not a natural object; it is a construction. Deciding what counts as a cheap company, over what universe, weighted how, rebalanced when, with which exclusions, involves a series of choices — each individually reasonable and each capable of being made otherwise.

Work on the most studied factor of all found that its standard construction rested on several apparently innocuous decisions that could have gone differently, and that the measured premium is smaller than first reported once alternatives are considered. The article on value investing covers that case in detail.

The consequence generalises. When a premium's size depends on construction choices, and the construction was chosen by someone who could see the result, the reported premium and the selection procedure are not independent — which is the multiple-testing problem arriving through a door nobody was watching.

What happens after publication

A factor that is real and known attracts capital, and capital moving toward a characteristic is precisely what would remove the premium attached to it. This is not a criticism of anyone; it is what a functioning market does with public information.

It produces an uncomfortable asymmetry for anyone evaluating a factor. Evidence from before publication is contaminated by the search that found it. Evidence from after publication is contaminated by the response to it. The clean test — forward from the moment of discovery, on a horizon long enough to matter — takes exactly as long as the horizon of the claim.