The Shanken Correction: Why Your Factor Model's t-Stats Are Too Good

This is a paper about a mistake that essentially everyone in empirical finance was making, quietly, for twenty years. It is not a glamorous paper. It has no memorable analogy, no famous chart, no Nobel Prize. What it has is a correction factor that, once you apply it, deflates a great deal of published confidence.

If you have ever run a Fama-MacBeth regression and reported a t-statistic, this paper is about you.

The problem: you cannot pretend an estimate is a fact

Here is the standard two-pass procedure that all of empirical asset pricing runs on:

Pass one. For each asset, run a time-series regression of its returns on the factors, and estimate its beta, its sensitivity to each factor.

Pass two. Take those estimated betas and use them as the explanatory variables in a cross-sectional regression to find out how much investors are paid per unit of each beta. This is the price of risk, and it is the number the whole exercise exists to produce.

The procedure is sound and it is what Fama and MacBeth taught the field to do. But look carefully at the seam between the two passes.

The betas that go into pass two are estimates. They were computed from a finite sample of noisy returns, and they carry sampling error. Some are too high, some too low, purely by chance.

And then pass two treats them as if they were exact, known numbers. It plugs them in as clean explanatory variables and computes standard errors as if there were no uncertainty in them at all.

That is the error. It is the classic errors-in-variables problem, and its consequences are not subtle. When your explanatory variable is measured with noise:

• The estimated slope is biased toward zero. Noise in the x-axis flattens the fitted line. Black, Jensen and Scholes had already worried about this, and grouping stocks into portfolios reduces it, but it does not eliminate it.
• The reported standard errors are too small. This is the more insidious problem, and it is the one Shanken solved. Because pass two ignores the uncertainty from pass one, it dramatically understates the true uncertainty in the final answer. Your t-statistics come out too large. Your p-values come out too small. Your result looks more significant than it is.

The key idea, via analogy

Suppose you want to know whether taller people earn more, but you do not have anyone's actual height. Instead, a friend eyeballed everyone from across a room and wrote down guesses.

You run your regression of income on the guessed heights and get a result. Now, how confident should you be?

Your statistical software will tell you exactly how much uncertainty comes from the scatter of incomes around the fitted line. What it will not tell you, because it does not know, is that your height data is itself made up of guesses. There is a whole second layer of uncertainty sitting underneath your analysis, entirely invisible to the software, and it makes the true error bars on your conclusion considerably wider than the ones on your screen.

Shanken's contribution is to work out, rigorously, exactly how much wider. He derives the extra term that accounts for the uncertainty in the first-stage betas propagating into the second-stage estimates, and it produces a multiplicative adjustment now universally known as the Shanken correction.

The correction has an intuitive structure. The size of the inflation depends on how large the estimated risk premium is relative to the volatility of the factor itself. In plain terms: if the factor is a strong, well-measured one whose premium is small relative to its own volatility, the correction is minor. But if you are claiming a large price of risk for a factor that does not move around much, then your betas are poorly pinned down relative to what you are asking them to explain, and the correction bites hard. Your impressive t-statistic shrinks.

Shanken also went further, tying together the traditional two-pass approach and the maximum likelihood methods economists preferred, showing how they relate, and extending the analysis to handle serial correlation in the factors.

Why it mattered

• It is now mandatory. Reporting Fama-MacBeth results without Shanken-corrected standard errors is, in a serious finance journal, a referee's first complaint. The correction is a standard column in essentially every asset pricing table.
• It deflated a lot of published findings. A number of results that had looked comfortably significant became marginal once the correction was applied. That is exactly what good econometrics is for.
• It is a general lesson in disguise. Any time you build a model in two stages and feed the output of stage one into stage two, you must carry the uncertainty forward. This mistake is everywhere: in machine learning pipelines that treat a fitted embedding as ground truth, in risk models that treat an estimated covariance matrix as known, in strategies that treat a fitted signal as a fact. Shanken's specific correction applies to beta-pricing models. His general warning applies to almost everything a quant does.
• It pairs naturally with GRS. Gibbons, Ross and Shanken had given the field a joint test of alphas. This paper gives the field honest standard errors on the risk premia themselves. Together they are the statistical backbone of factor model evaluation.

The honest limitations

• It fixes the standard errors, not the bias. The correction gives you honest uncertainty around your estimate. It does not remove the attenuation that pushes the estimate itself toward zero. Portfolio grouping and instrumental variable approaches address that, imperfectly.
• It assumes a specific error structure. The derivation relies on assumptions about how returns are distributed, conditionally homoskedastic and well-behaved. Real returns have clustered volatility and fat tails, so the corrected standard errors are better but still not exact.
• It does not save you from a wrong model. Correct standard errors on a misspecified model are correct standard errors on nonsense. Shanken tells you how much you should trust your number, not whether you were computing the right number.
• It cannot fix data mining. If you have searched over hundreds of factors to find the ones that price well, no single-test correction, Shanken's or anyone else's, restores the honesty of your t-statistic. That is a separate problem, and one that Harvey, Liu and Zhu would later make painfully clear.
• The correction is often small in practice, which invites complacency. For well-behaved factors with plenty of data it is a modest adjustment, which tempts people to skip it. The times it matters most are precisely the times when a marginal result is being pushed as significant, which is when skipping it is least excusable.

The one-line takeaway

Shanken showed that because factor tests use estimated betas as if they were known facts, the resulting t-statistics are systematically too flattering, and he derived the correction that finance now applies as a matter of course to keep itself honest.