Taming the Factor Zoo: Is Your New Factor Actually New?

Suppose you have discovered a new stock market factor. You want to publish it. What must you prove?

The obvious answer is "that it predicts returns." But that is not nearly enough, and the reason is the factor zoo: hundreds of factors have already been published. The bar is not "does your factor work." The bar is "does your factor add anything that the hundreds of existing factors do not already provide?"

That turns out to be a genuinely hard statistical problem, and Guanhao Feng, Stefano Giglio and Dacheng Xiu are the ones who solved it properly.

The problem: you cannot control for everything at once

The standard way to show a factor is new is to control for the known factors. Run your new factor alongside the market, size, value, momentum and profitability, and show it still has explanatory power.

But which controls do you use? Choose the three-factor model and you are ignoring a hundred other published factors. Choose all of them and you have a statistical disaster: you cannot run a regression with hundreds of highly correlated variables and a few hundred data points. The estimates fall apart.

So researchers pick a handful of controls. And here is the corrosive incentive: the researcher gets to pick which factors to control for. Naturally, one tends to pick the controls against which one's own factor looks best. It need not be dishonest. It is just what happens when the choice is free and the outcome depends on it.

The alternative, throwing every factor into a machine learning variable selector and keeping whichever survive, sounds appealing but is statistically wrong, and this is the subtle heart of the paper. Selection methods make mistakes. If the selector accidentally drops a control factor that genuinely mattered and that happens to be correlated with your new factor, then your new factor absorbs the dropped factor's explanatory power and looks significant when it is not. This is classic omitted variable bias, sneaking in through the back door of an automated procedure. A naive machine learning approach does not just fail to solve the problem, it manufactures false discoveries.

The key idea via analogy: hire the new player only if the team gets better

Think of an asset pricing model as a sports team. Hundreds of players (factors) have been proposed. You cannot field them all. Somebody arrives with a new player and claims they are essential.

The right test is not "is this player any good in isolation?" Almost everyone is good in isolation. The right test is: given the players we already have, does adding this one make the team measurably better?

And the trap is that you have to be careful about who is currently on the field. If you evaluate the new recruit against a weakened team, they will look like a superstar. If you accidentally bench someone who plays the same position, the recruit looks essential when they are merely a substitute.

Feng, Giglio and Xiu's contribution is a procedure that handles this honestly. The technical machinery is a double-selection approach, borrowed from modern statistics: rather than selecting variables once and hoping, they run selection twice, once to find which existing factors matter for explaining returns and once to find which existing factors are correlated with the new candidate factor. Keeping the union of both sets is what protects against the omitted variable bias that dooms the naive approach. Only then do they ask whether the new factor still contributes.

The result is a disciplined, automatic answer to the question "is this factor new?", with proper statistical guarantees, and with no scope for a researcher to cherry-pick flattering controls.

The finding

They applied it to a large collection of published factors. The verdict:

Most of the new factors were redundant. Given the hundreds of factors already in the literature, the typical new arrival added nothing that existing ones did not already capture. It was a substitute, not a discovery.

A few genuinely survived. Some factors did carry statistically significant explanatory power beyond everything that came before. The zoo is not entirely populated by clones. But the survivors are a small minority.

Why it mattered

• It set a proper standard for a new factor. Before this, "controlling for known factors" was a vague and manipulable exercise. Afterwards there was a defensible procedure, and the burden of proof shifted decisively onto anyone proposing a new factor.
• It brought modern statistics to asset pricing responsibly. This is not machine learning sprayed at a finance problem. It is a careful use of high-dimensional selection methods that explicitly corrects for the bias those methods introduce. That combination is why it won a place in the top journal in the field.
• It complements the other zoo papers. Harvey, Liu and Zhu said the significance bar is too low. Hou, Xue and Zhang said the empirics were sloppy. Feng, Giglio and Xiu say that even a factor with clean empirics and a high t-statistic may be redundant, which is a third and quite different way to be worthless.
• It is directly useful to practitioners. Anyone maintaining a library of signals faces exactly this question: is this new signal adding information, or am I just paying to trade the same bet twice? The method answers it.

The honest limitations

• Redundant is not the same as wrong. A factor can be entirely real and still be redundant, in the sense that other factors already span it. That matters for asset pricing theory but may matter less to a trader who finds the new version cheaper or faster to trade.
• The answer depends on the zoo you feed it. The test asks whether a factor adds value beyond a given set of existing factors. Change the set and the verdict can change. The method removes the researcher's discretion over controls, but somebody still had to assemble the factor library.
• It inherits the data problems of everything else. If many of the factors in the library are themselves noise, then testing a new factor against them is testing against a wall of noise, which is not obviously the right benchmark.
• It cannot resurrect economics. The test is purely statistical. A factor with a compelling economic story and marginal statistical contribution may deserve more respect than the method affords it.

The one-line takeaway

Feng, Giglio and Xiu built the first statistically honest way to ask whether a new factor adds anything beyond the hundreds already published, carefully avoiding the bias that naive machine learning introduces, and concluded that most new factors are redundant while a small handful genuinely are not.