Lesson 3 of 7
The size premium, re-examined
Run the small-company test on 4,383 US stocks over three years and the median says the opposite of the mean. Both are computed from the same rows.
By Javier Sanz
Founder of ValueScreener. Built Ninety Nine, a securities services provider, and sold it to a tier-one global bank; ran operations and internationalization at Alpaca, the brokerage API.
Who writes thisEditorial processLinkedIn (opens in a new tab)X (opens in a new tab)
In one minute
- Adding a premium to a discount rate for small size is common practice, and the evidence behind it is contested rather than settled.
- On 4,383 US stocks over one three-year window, the median smallest-quintile return was minus 12.9 percent and the mean was plus 66.4 percent.
- A test this data cannot pass is still worth running, because knowing why it fails is what stops the number being used.
A size premium is one of the few adjustments analysts make to a discount rate without a procedure. It is worth asking what the evidence for it looks like when you try to produce some.
Three-year price change by size quintile, ranked at the start of the window
- Smallest quintile, median−12.9%
- Second, median−1.4%
- Third, median9.5%
- Fourth, median13.4%
- Largest quintile, median21.1%
The result, both ways
| Quintile at the start | Median change | Mean change |
|---|---|---|
| Smallest | -12.9% | +66.4% |
| Second | -1.4% | +46.9% |
| Third | +9.5% | +46.1% |
| Fourth | +13.4% | +38.5% |
| Largest | +21.1% | +41.5% |
On the median the ranking is monotonic and runs against a small-company premium: the typical small company lost money while the typical large one gained. On the mean the smallest quintile wins by 25 points.
Both columns are computed from the same rows. The gap between them is a handful of very large winners in the small bucket, which lift an average and cannot lift a median. If a strategy holds every name in the quintile, the mean is the relevant figure; if it holds a few, the median is much closer to the experience.
Why this test does not settle anything
Four objections, and each one is larger than the effect being measured.
The failures are missing. Only companies still listed today are in the sample. Everything that was delisted, acquired at a loss or wound up between 2023 and 2026 is absent, and those are concentrated in the small end. The true small-company numbers are worse than the table.
It is price only. No dividends, which understates the large end, where more of the return arrives as income.
One window. Three years is a sample of one for a claim about long-run expected returns.
The ranking is an estimate. Start-of-window size here is today's market value scaled by the price change, which assumes the share count did not move. A company that issued heavily is misplaced.
What the literature actually supports
The 1992 Fama and French paper is where size entered most practitioners' vocabulary, as one of two variables that outperformed market beta in explaining the cross-section of US returns over 1963 to 1990. Its status since has been argued over rather than confirmed, and the general finding that published return predictors weaken after publication applies here as much as anywhere: McLean and Pontiff, in the Journal of Finance in 2016, measured exactly that decay across a large set of documented predictors.
The honest position for a discount rate is therefore narrow. If you add a size premium, say how large, where the number came from, and that the effect it represents is contested. If you cannot do that, the addition has no procedure behind it, and an unlabelled point on a discount rate is worth about a fifth of the valuation.
Check yourself
The smallest quintile had a median return of minus 12.9 percent and a mean of plus 66.4 percent. Why?
- One of the two was computed wrongly
- A few very large winners lift the mean and cannot move the median
- The quintiles were formed at different dates
Show the answer
Answer: B.
Small-company returns are heavily skewed. Which average you report decides which direction the result appears to run.What is the largest problem with the test above?
- Three years is too short a window
- The sample contains only companies still listed today, so the failures are missing and the small end is flattered
- It uses medians rather than means
Show the answer
Answer: B.
Survivorship removes exactly the observations the small end needs. The true figures are worse than the table, by an amount this data cannot measure.What does a defensible size adjustment to a discount rate require?
- A round number, such as one point
- A stated size, a named source for it, and an acknowledgement that the effect is contested
- Nothing: the effect is well established
Show the answer
Answer: B.
An unlabelled addition to a rate is untraceable later and moves the answer by more than most of the analysis above it.Sources
- Computed for this lesson from the product's own stored daily closes for 4,383 US-listed common equities, early September 2023 to 2026-09-15.
- Fama and French, Journal of Finance, 1992: size and book to market against market beta over 1963 to 1990.
- McLean and Pontiff, Journal of Finance, 2016: the decay in documented return predictors after the papers describing them are published.
What this leaves out
- The sample holds only companies still listed on 2026-09-16, so it is biased upward by an unmeasured amount, most of all in the small quintile.
- Returns are price changes with no dividends, over a single three-year window, with no adjustment for risk of any kind.
- Start-of-window size is estimated by scaling today's market value by the price change, which assumes a constant share count.
- Nothing here is advice about any security.