Key takeaways
- The Fama/French size factor compounded at 1.44% a year from July 1926 to June 2026, with a t-statistic of 1.84 on its monthly mean.
- Banz published in 1981. From April 1981 to June 2026 the factor compounded at -0.30% a year, a cumulative -12.72% over 45 years.
- The factor peaked in July 1983 and has never regained that level. At June 2026 it stood 31.16% below it, after 515 months underwater.
- Since 1926 the factor has averaged 1.91% in January and 0.01% across the other 1,100 months. Compound the non-January months alone and you get -33.16%.
- The smallest decile of US listed stocks holds 38.9% of the companies and 0.35% of their market value. Since 1981 it has trailed the largest decile by 1.93 points a year.
The oldest documented anomaly has the weakest record since
Do small companies still earn more than large ones? The size effect was the first anomaly anyone documented, so the question has an unusually long public answer.
Rolf Banz published the finding in the March 1981 issue of the Journal of Financial Economics. Here is what the data has done since, taken from the factor series Fama and French give away. Their size factor — the return on small stocks minus the return on big ones — compounded at 1.44% a year from July 1926 to June 2026. From April 1981, the month after Banz appeared in print, to June 2026, it compounded at -0.30% a year. That is a cumulative -12.72% across 45 years.
Not a smaller premium. No premium at all.
The t-statistic on the full series is 1.84. That number weighs the average against how much it jumps around month to month, and anything under about two is weak evidence. Across the 45 years since publication the same statistic is 0.14. Over exactly those months the US market factor compounded at 7.75% a year and the value factor at 2.61%. Equities did fine. Small simply didn't beat big.
What Banz found in 1981, and the warning he attached to it
Banz sorted New York Stock Exchange stocks by market value and measured what each group earned after adjusting for market risk. One number from that paper travelled further than the rest: "The average excess return from holding very small firms long and very large firms short is, on average, 1.52 percent per month or 19.8 percent on an annualized basis." His test window was January 1936 to December 1975.
That 19.8% a year isn't the size factor. Banz built equally weighted portfolios of ten, twenty or fifty of the very smallest firms, levered them so the portfolios shared the same market beta, and held them five years with monthly rebalancing. Run French's broad size factor over the same 1936 to 1975 window and it compounded at 1.41% a year.
Banz made the concentration point himself. Of the size-sorted portfolios he wrote: "The smallest firms have, on average, very large unexplained mean returns. There is no significant difference between the residuals of the remaining portfolios." The whole effect sat in the tiniest names, and he said so on the page.
He attached the practical warning in the same paragraph as the headline figure. The strategy "leaves the investor with a poorly diversified portfolio", and his own table showed it "would not have been successful in every five year subperiod". If you have read how many stocks it takes to be diversified, a ten-stock micro-cap book is not a technicality.
Fama and French then made size a standard factor. In their 1992 paper, "Average returns fall from 1.64% per month for the smallest ME portfolio to 0.90% for the largest", measured from 1963 to 1990.
The premium by decade, compounded from the monthly file
The chart above takes the monthly size factor and compounds it inside each calendar decade. The source is the Fama/French three-factor file in Kenneth French's data library, built from the June 2026 CRSP database, so these are not estimates of the series. They are the series.
Left to right: 8.75% a year in the 1930s, 4.50% in the 1940s, -0.86% in the 1950s, 4.45% in the 1960s, 2.81% in the 1970s, -0.38% in the 1980s, -2.06% in the 1990s, 4.21% in the 2000s, -0.50% in the 2010s, and -2.20% a year so far in the 2020s.
Five of the ten decades are negative, and four of those five fall after Banz published. Of the three classic factors in this library, size is the only one whose premium went to zero. Value shrank but stayed positive, which is the subject of the value premium's own post-publication record, and momentum kept a large gross premium and gave much of it back in turnover. Both were computed from the same library on the same conventions.
Treat the table gently. Ten decades is ten observations, and the boundaries are an accident of the calendar rather than a feature of markets.
Start the out-of-sample clock in 1976 and you get 1.21%, not zero
There are two defensible dates for "out of sample", and they don't agree.
Banz's sample ended in December 1975. Start the clock there, which is what AQR's researchers do, and the factor compounded at 1.21% a year to June 2026, with a t-statistic of 1.20. Start it at publication in 1981 and you get the -0.30% above. That is a real gap, and the friendlier number deserves to be on the page.
One stretch explains all of it. From January 1976 to December 1986 the factor compounded at 7.08% a year, with a t-statistic of 2.80 — nearly five times its own hundred-year average. From January 1987 to June 2026 it compounded at -0.37%.
So the size premium's strongest modern stretch sits in the gap between the end of Banz's evidence and the point where enough money could act on the paper. That is suggestive rather than conclusive, and it is exactly the sort of coincidence the data-mining critique is built on.
Forty-three years below the 1983 high
Averages hide the shape of the thing. Run the factor as a cumulative index and it peaked in July 1983, two years after publication. It bottomed in March 1999, 54.95% lower.
Then it didn't recover. At June 2026 the index was still 31.16% below its July 1983 level, having spent every one of the 515 months since underwater. A drawdown that deep and that long outlives the career of anyone who started the position, which is the practical gap between a factor's average and a person's experience of holding it — and why drawdown and volatility measure different risks.
Almost the entire premium happens in January
Here is the finding that makes the rest of the record hard to read. Since July 1926 the size factor has averaged 1.91% in January, with a t-statistic of 5.52. Across the other 1,100 months it averaged 0.01%, t-statistic 0.10.
Compound the Januarys on their own and the total is 524.18%. Compound every other month and the total is -33.16%. One month of the year carries the century, and the remaining eleven lose money together.
That result is not ours. Asness, Frazzini, Israel, Moskowitz and Pedersen report it in the Journal of Financial Economics: "The returns to SMB are enormous in January at 2.3% for the month and the 1–10 spread in size decile returns is even larger at 6.8% in January. However, from February through December, SMB delivers a negligible 4 bps and the 1–10 portfolio spread is −1 bp, both of which are statistically and economically zero." Their sample ends in December 2012; ours runs to June 2026 and finds the same shape.
Alquist, Israel and Moskowitz put it more bluntly in their 2018 survey: "The bottom line is that other than in January, there is not and never was a size premium." Over 1926 to 2017 they measure January at "2.1% return in the month of January alone, with a t-statistic of 6.25", against non-January months showing "literally zero size premium (average return of 0.0% from 1926 to 2017)".
The seasonal has faded along with everything else. Since April 1981 the factor's January average is 0.42%, with a t-statistic of 0.89. The other months average -0.02%. In the same survey, January over 1976 to 2017 "averages only 1.0% per month in this period compared to the 2.1% return it exhibited in January over the longer sample dating back to 1926, with a t-statistic of 2.39".
The premium lives in 0.35% of the market by value
Now the composition problem, in numbers from French's size-portfolio file for June 2026. The smallest decile of US listed stocks held 1,242 of the 3,192 companies across the ten deciles — 38.9% of the names — and 0.35% of their combined market value. The largest decile held 78.22% of the value.
That lopsidedness is by construction: "The portfolios are constructed at the end of each June using the June market equity and NYSE breakpoints." Breakpoints from the New York Stock Exchange, applied to every listed US stock, put the whole micro-cap tail into one bucket.
Weighting then decides most of the answer. Over the full century the smallest decile returned 12.15% a year value-weighted and 15.95% a year equally weighted. The largest decile returned 10.05%. So the spread between smallest and largest is 2.10 points a year on market weights and 6.26 points on equal weights. Most of the famous version of the size premium is a weighting decision, not a return earned by being small.
Remove the tiniest names and what is left stops paying
The obvious test is to drop the bottom decile and look at the next one up. Over the full sample the second-smallest decile beat the largest by 1.39 points a year on market weights, against 2.10 points for the smallest. Since 1981 both are negative: the smallest decile trailed the largest by 1.93 points a year, the second-smallest by 0.84.
The published work reaches the same place with different data. AQR's survey records that "Horowitz, Loughran, and Savin (2000) found that removing stocks with less than $5 million in market cap eliminates the small-firm premium", and that the risk-adjusted premium "declines rapidly when we remove the smallest 5% of firms as well, which corresponds to firms with a market cap of about $18.8 million".
Costs then finish what is left. On their estimates, trading the size factor "at $200 million incurs about 88 bps per year; at $2 billion, about 152 bps per year; and at $5 billion, 240 bps per year". Set that against a gross spread of about two points a year that has been negative for four decades, and the arithmetic doesn't leave much room.
The strongest counter-case: size works once you control for junk
The best defence of size comes from some of the same authors as the sceptical survey, which is worth knowing before you weigh it.
Asness, Frazzini, Israel, Moskowitz and Pedersen argue the size effect has been measured into a headwind. Small firms are, on average, low-quality firms — less profitable, more volatile, more indebted. Quality carries a premium of its own, so a plain small-minus-big portfolio is long exactly the junk that a quality strategy is short. Strip that out and something different appears: "SMB loads significantly negatively on quality, driving the SMB alpha from 14 to 49 bps per month, which is almost five standard errors from zero (t-statistic = 4.89)."
Their abstract claims the whole objection list falls with it — a premium that is "stable through time, robust to specification, not concentrated in microcaps, more consistent across seasons, and evident for non-price-based measures of size, and these results hold in 30 different industries and 24 international equity markets".
They are candid about the raw factor. Over their sample, plain size delivered "55 bps per month, with a t-statistic of 2.32", while value and momentum posted t-statistics of "3.7 and 4.6" across the same months.
Refereeing it: the resurrected premium is a different portfolio
The two camps agree about the record. What they disagree about is which portfolio deserves the name.
The quality-controlled result is a genuine finding, and it is the best case anyone has for size. It is also not the trade Banz described, and not the series in the data library. It is a hedged position: long small, short big, and short junk in whatever size the regression demands. Somebody holding a small-cap index fund holds the unhedged version, and the unhedged version has compounded at -0.30% a year since 1981.
The critics have a soft spot too. Their sharpest exhibit is January, and nobody has a settled explanation for it. Tax-loss selling, window dressing and turn-of-year cash flows are all proposed, and none is agreed. An unexplained pattern is thin ground for the claim that the premium was never real.
The delisting-bias argument has the same shape. Prices for stocks that vanished were missing from the old databases, and Shumway's backfill found "the delisting return is -55%, on average, for NASDAQ firms and -30%, on average, for NYSE/Amex firms when the delisting is for performance-related reasons". Small firms delist more often, so the fix shrank the original numbers. That damages Banz's estimate. It says nothing about the past four decades, which were measured on the corrected data.
Outside the US the record is worse, not better
Value's strongest defence is that it kept working outside America. Size has no such defence.
French's developed-markets-ex-US size factor starts in July 1990, so nearly all of it postdates Banz. To June 2026 it compounded at -0.77% a year, a cumulative -24.17%, with a t-statistic of -0.47. The value factor in that same file compounded at 4.70% a year over the same months, t-statistic 3.65. Same countries, same construction, opposite outcomes.
The academic work agrees. Across 23 non-US markets from 1986 to 2012, Asness and co-authors found size "averaging 13 bps with an insignificant t-statistic of 0.91". How much of your equity sits abroad decides which of these records you actually lived through, which is one reason the foreign share of your equities is a bigger decision than it looks.
What this data cannot tell you
Four limitations, and each one bites.
The factor is one definition of size, built from six value-weighted portfolios sorted on size and book-to-market and rebalanced each June. Change the weighting, the universe or the rebalancing date and you get a different series — the gap between the 2.10 point value-weighted spread and the 6.26 point equal-weighted one is the size of that sensitivity.
It is also a costless paper portfolio. It pays no spread, no borrow fee on the short leg, no commission and no tax. Every after-cost figure quoted here comes from a separate study with its own cost model.
The January result is a pattern without an accepted cause, which cuts both ways. It undermines the premium and it undermines the confidence of anyone using it as proof.
And a century of monthly US returns is one history rather than a distribution of histories. This is a backtest, not a forecast, and it contains exactly one publication event to learn from. Drawing a general law about anomalies decaying from a single case is not something the sample supports.
What would change the conclusion
A decade of size paying outside January. The non-January average since 1926 is 0.01% a month. If that turned reliably positive while the January spike stayed dead, the seasonality argument would lose its force, and the case that small stocks earn something would rest on evidence it has never had.
A live record for the quality-controlled version. The 49 basis points a month emerges inside a regression that shorts junk. A fund running that trade with real costs, real borrow and real capacity would settle the argument in a way another backtest cannot.
The ex-US series turning up. At -0.77% a year since 1990, the international evidence is currently doing the most damage. If it converged on the value factor's 4.70%, the data-mining reading of the whole size literature would get much harder to hold.
The figure worth tracking is not the annual return. It is the February-to-December average, because that is the one number on which the two explanations make opposite predictions. A risk premium should be paid in every month. A turn-of-year liquidity artefact should be paid in one — and for a hundred years, it has been.