Factor Investing: What the Premiums Are and Why They Shrink

Factor investing is the practice of tilting a portfolio toward stock characteristics that research has linked to higher average returns, such as small size, low price relative to book value, or strong recent performance. The documented premiums are averages over long periods, not schedules: in the Kenneth French Data Library figures analyzed below, the size, value and momentum premiums averaged noticeably less in 1991 to 2025 than in the decades before, and the compounded monthly value spread went through a drawdown of nearly 58% that took until September 2020 to bottom out.

Key takeaways

• A factor premium is the average return gap between two groups of stocks sorted on a characteristic. It is a statistical regularity in past data, not a fund you can buy or a return you are owed.
• In annual data from the French library, the average premium fell from 3.59% to 0.51% (size), 5.44% to 1.99% (value) and 10.68% to 4.75% (momentum) when 1991 to 2025 is compared with 1927 to 1990. Profitability went the other way, from 2.49% in 1964 to 1990 to 4.06% in 1991 to 2025, so decay is not uniform.
• A peer-reviewed study of 97 published predictors found portfolio returns 26% lower out of sample and 58% lower after publication; a separate paper argues new factors should clear a t-ratio of about 3.0, not 2.0.
• Value's monthly spread, compounded, fell about 58% from December 2006 to September 2020 and, through August 2026, had not regained its 2006 peak. Holding a factor means accepting dry spells that can last longer than many investors' patience.
• Factor funds typically cost more than cap-weighted index funds, partly because their indexes trade more often, and backtests do not predict future results.

What a factor is, and the main ones

A factor, in this literature, is a characteristic shared by a group of stocks. A factor premium is the average return of stocks with one value of the characteristic minus the average return of stocks with the opposite value. The standard public series come from the Kenneth French Data Library, which builds them from US stocks listed on the NYSE, AMEX and NASDAQ and publishes monthly and annual values.

Market

The market factor is the oldest. It is rooted in the Capital Asset Pricing Model, set out in William Sharpe's 1964 Journal of Finance paper among others, and it is measured in the French files as the market return minus the one-month Treasury bill rate. Later work questioned beta as the sole driver of average returns: the abstract of Fama and French's 1992 paper reports that firm size and book-to-market equity together explained differences in average returns, and that when beta varied independently of size, the relation between beta and average return was flat.

Size and value

Fama and French's 1993 paper in the Journal of Financial Economics established the three-factor model of market, size and value. According to the library's definitions, SMB (small minus big) is the average return of three small-company portfolios minus that of three big-company portfolios, and HML (high minus low) is the average return of two value portfolios minus that of two growth portfolios, sorted on book-to-market.

Momentum

The 1993 Jegadeesh and Titman paper reported that buying past winners and selling past losers produced significant positive returns over three- to twelve-month holding periods, and that the profits were not explained by systematic risk. Carhart's 1997 study found that much of the persistence in mutual fund returns traces to that one-year momentum effect, while individual funds did not earn higher returns by following the strategy. The French momentum series (Mom) is the average of the big and small high prior-return portfolios minus the average of the low ones, with prior return measured from month -12 to month -2.

Profitability and investment

The 2015 five-factor model adds two factors to size and value. Per the library's definitions, RMW (robust minus weak) is the return spread between firms with robust and weak operating profitability, and CMA (conservative minus aggressive) is the spread between low-investment and high-investment firms. The abstract adds that, in the authors' sample, the value factor becomes redundant once profitability and investment are included.

Low volatility and quality

Two further factors recur in fund marketing. In "Betting Against Beta", Frazzini and Pedersen report that a factor going long leveraged low-beta assets and short high-beta assets earned significant risk-adjusted returns across several asset classes. In "Quality Minus Junk", Asness, Frazzini and Pedersen define quality as profitability, growth and safety and report a long-short quality factor with significant risk-adjusted returns in the US and 24 other countries.

Why a premium might exist

Researchers disagree about the reason, and the answer bears on whether a premium can last.

The risk-based view says the premium pays investors for bearing a risk that the market does not otherwise capture. Fama and French's 1996 paper found that, except for short-term return continuation, many return anomalies disappeared in a three-factor model, a result they said fits rational pricing models, while also weighing irrational pricing and data problems as possible explanations.

The behavioral view says the premium reflects investor error. Lakonishok, Shleifer and Vishny (1994) argued that value strategies earn higher returns because they exploit the suboptimal behavior of the typical investor, not because they are fundamentally riskier.

The two views predict different futures. A risk premium should persist as long as the risk does. A mispricing should shrink as more investors learn of it and trade against it, which is the pattern tested in the next sections.

What the French data show

The table compares simple arithmetic averages of calendar-year factor returns in two windows. It uses the annual files from the Kenneth French Data Library, downloaded on 2026-10-20: the three-factor file for market, SMB and HML, the momentum file for Mom, and the five-factor file for RMW and CMA. The files state they were created with the 202608 CRSP database, and the annual series end in 2025. The five-factor annual series starts in 1964, so RMW and CMA use 1964 to 1990 as the earlier window. The t-ratio is the average divided by its standard error, computed from the annual values.

Four points follow from the numbers. First, size, value, momentum and investment averaged less in the later window, and none of those later-window t-ratios reaches 2.0. Second, the market premium did not shrink, so the pattern is not a general fading of stock returns. Third, profitability went the other way, from 2.49% to 4.06%, so the decay is not universal. Fourth, the series are volatile: the standard deviation of annual HML was 17.6 percentage points in 1991 to 2025, which is why an average of 1.99% is hard to distinguish from zero over 35 years.

These are return spreads computed from the returns of the stocks in each portfolio, not results of a fund. Built that way, they carry no fund fees, short-selling costs or trading costs, and an investor who traded them would have faced all three. Novy-Marx and Velikov (Review of Financial Studies, 2016), who studied anomaly strategies after trading costs, report that "transaction costs always reduce strategy profitability." The 1991 split is a convenience, not a finding: the data cannot say whether the decline reflects publication, chance, a changing risk environment or something else.

Why premiums shrink once known

The most-cited test is McLean and Pontiff (Journal of Finance, February 2016). They studied 97 variables shown to predict cross-sectional stock returns and found portfolio returns 26% lower out of sample and 58% lower after publication. They treat the 26% as an upper bound on data-mining effects and attribute the remaining 32 percentage points (58% minus 26%) to publication-informed trading. Declines were greater for predictors with higher in-sample returns, and the published-predictor portfolios became more correlated with one another after publication. Their reading is that investors learn about mispricing from academic papers.

The factor zoo and the t-ratio hurdle

A second problem is multiple testing. If hundreds of researchers test hundreds of characteristics, some will look significant by chance. In Harvey, Liu and Zhu's paper (Review of Financial Studies, 2016), the authors build a multiple-testing framework and conclude that a newly discovered factor should clear a t-ratio of about 3.0, not the conventional 2.0. They also argue that most claimed research findings in financial economics are likely false.

Illustrative example. Suppose a researcher tests 20 characteristics, none of which truly predicts returns, and the tests are independent. At the 2.0 hurdle each test has about a 4.6% chance of looking significant by luck, so the researcher expects 0.9 false discoveries and has about a 61% chance of finding at least one. At the 3.0 hurdle the per-test chance falls to about 0.27%, and the chance of at least one false discovery among the 20 drops to about 5%, roughly 1 set in 19. These figures use a normal approximation and are not how Harvey, Liu and Zhu derive their threshold, which comes from a fuller framework. They show why a single t-ratio of 2.2 found in a large search deserves less trust than the same number from one pre-specified test.

A dry spell: value, 2007 to 2020

Using the monthly HML series from the same library (through August 2026), and compounding it as an illustration of the spread rather than a tradable return, the series peaked in December 2006 and fell 57.8% to a low in September 2020, a span of 13 years and 9 months. Compounded across January 2007 to December 2020, the monthly spread lost 55.8%. Through August 2026 the series stood at about 68% of its December 2006 level, so it had not regained the old peak.

The annual figures tell the same story from another angle. HML averaged -6.70% per year across 2007 to 2020 and was negative in 10 of those 14 years. Across all 99 years from 1927 to 2025, it was negative in 40. From 2021 to 2025 it averaged +7.52% per year, with negative years in 2023 and 2024, which shows how mixed the record can be inside a few years.

The lesson is about holding periods. A premium with a positive long-run average can still trail for longer than a typical investment plan, and the data alone do not say whether 2007 to 2020 was a rough patch in a lasting premium or evidence that the premium had faded.

Costs and access through funds

Investors typically reach factors through exchange-traded funds that FINRA calls smart beta products. According to its investor insight dated August 9, 2022, smart beta indexes may weight stocks equally, by fundamentals such as revenues or dividends, or by volatility. FINRA says these products usually cost less than actively managed funds but more than funds tracking market-cap-weighted indexes, and that indexes rebalance more often, which "can increase investor costs."

FINRA lists further risks: less diversification, and exposure that can lean toward a particular sector, country or company size; thin trading and wide bid-ask spreads in newer products; and returns that may differ significantly from a cap-weighted index. It cautions that back-tested results neither predict future performance nor perfectly replicate past performance.

Frequently asked questions

Is smart beta the same as factor investing?

Not exactly. FINRA describes smart beta as a set of indexes that depart from market-cap weighting, and equal weighting is one of its examples; FINRA notes it overweights smaller companies relative to a cap-weighted index. Equal weighting is a weighting rule, not a sort on a characteristic such as book-to-market, so a smart beta label alone does not say which factor, if any, a product targets. FINRA's first suggested question is what the product's strategy is, answered in the prospectus and the index provider's methodology.

Can an investor buy HML or SMB directly?

No. The library builds them from six value-weight portfolios formed on size and book-to-market, and the factor is the difference between portfolio groups, so it involves a long and a short side. A fund with a value label holds its own selection and weighting rules, which is why FINRA advises reading the index methodology.

Is the shrinkage just an artifact of picking 1991 as the split?

Not in these files. Splitting the annual series at 1981 instead, the averages before and after are 4.64% and -0.07% for SMB, 4.85% and 3.47% for HML, and 11.05% and 5.63% for Mom. Splitting at 2001, they are 3.00% and 1.02%, 5.58% and 0.20%, and 11.30% and 0.54%. Splits at 1986 and 1996 show the same direction for all three. The test shows the decline is not tied to one chosen year; it cannot show why the decline happened.

How often are factor portfolios rebalanced?

The data library says momentum and short-term reversal portfolios are reconstituted monthly, and the other research portfolios annually. More frequent rebalancing means more trading in any fund that copies the approach, which is the turnover cost FINRA flags.

Did the market premium shrink too?

Not in these data. The later-window average of annual market returns above T-bills was higher at every split tested: 9.18% against 8.60% splitting at 1981, 9.68% against 8.31% at 1986, 10.36% against 8.04% at 1991, 9.73% against 8.49% at 1996, and 9.09% against 8.79% at 2001. That does not say the market premium is stable going forward, only that the shrinkage seen in size, value and momentum is not a general feature of the series.

Questions to ask before a tilt

1. Which definition is the fund using? "Value" can mean book-to-market, earnings, dividends or a blend, and each gives a different portfolio.
2. What is the evidence window? Check whether the premium appears after the paper that documented it, not only before.
3. What does the fund charge, and how often does its index trade? Compare the expense ratio and turnover with a plain cap-weighted index fund.
4. Could you hold through a 13-year drawdown like value's 2006 to 2020 spell without changing course? If not, the premium may never reach you.
5. Is the claim a backtest? Treat it as a historical illustration, not a forecast.

Sources

• Kenneth R. French Data Library (annual and monthly factor files, CRSP 202608 vintage; definitions at the 3-factor and 5-factor pages)
• McLean and Pontiff, "Does Academic Research Destroy Stock Return Predictability?", Journal of Finance, 2016
• Harvey, Liu and Zhu, "... and the Cross-Section of Expected Returns", Review of Financial Studies, 2016 (NBER w20592)
• Fama and French (1993), Journal of Financial Economics and (1992), Journal of Finance
• Fama and French (2015), "A five-factor asset pricing model"
• Jegadeesh and Titman (1993) and Carhart (1997)
• Novy-Marx and Velikov (2016), "A Taxonomy of Anomalies and Their Trading Costs", Review of Financial Studies
• FINRA, "Smart Beta: What You Need to Know" (August 9, 2022)