Factors and the evidence

Chapter 7 noted CAPM's strong clause: once beta is known, no other characteristic affects expected return. This chapter is what happened to that clause.

What a factor is

A factor is a characteristic that sorts securities into groups whose returns differ systematically. It is measured as the return on a long-short portfolio: buy the securities with the characteristic, sell the ones without, and the difference is the factor's return.

Ken French's own definitions, for the two oldest:

SMB (Small Minus Big) is the average return on the three small portfolios minus the average return on the three big portfolios

HML (High Minus Low) is the average return on the two value portfolios minus the average return on the two growth portfolios

Reading those carefully tells you what a factor return is not. It is not the return on small companies. It is the return on small minus big — a self-financing spread, with roughly zero net market exposure by construction. So a factor premium is not available by simply buying the cheap things, and the gap between the spread and what an investor can hold is a large part of this chapter.

What they earned

The factors are published monthly from July 1926. Over the century to August 2026:

Factor	Annualised mean	Annualised sd	$t$-statistic	Sharpe
Market minus riskless	8.35%	18.37%	4.55	0.455
Small minus big (SMB)	1.99%	10.90%	1.83	0.183
Value minus growth (HML)	4.22%	12.31%	3.43	0.343

The market row is the control. 8.35% a year with a $t$ of 4.55 — this is the premium CAPM says exists, and it does. Its Sharpe ratio of 0.455 is close to the 0.40 that chapter 9 took from Sharpe's own paper for a developed market, which is a reassuring consistency check across two unrelated sources.

The other two rows are the problem for CAPM. Size and value are characteristics the model says should not be priced once beta is accounted for. Over a century, value's spread earned 4.22% a year with a $t$-statistic of 3.43. That is not a rounding error and it is not a short sample.

And then the premia fell

Fama and French's papers establishing size and value appeared in 1992 and 1993. Splitting the data there:

Factor	1926–1992	1993–2026	Last 20 years
Market minus riskless	7.92% ($t$ = 3.27)	9.19% ($t$ = 3.50)	10.56% ($t$ = 3.00)
Small minus big	2.75% ($t$ = 2.05)	0.48% ($t$ = 0.26)	−0.47%
Value minus growth	5.26% ($t$ = 3.37)	2.17% ($t$ = 1.10)	−1.16%

Read the three rows against each other, because the comparison is the finding.

The market premium did not decay. It was 7.92% before 1993 and 9.19% after, and 10.56% over the last twenty years. Whatever happened to the other two did not happen to it.

Size went to nothing. 2.75% becomes 0.48%, with a $t$-statistic of 0.26 — indistinguishable from zero.

Value more than halved and then went negative. 5.26% becomes 2.17% with a $t$ of 1.10, and the most recent twenty years are −1.16% a year.

This is the same shape the Technical analysis subject found in a completely different literature: rules that worked in the data they were discovered in, and stopped working afterwards. The mechanism proposed there — a published profitable rule attracts users, who compete it away — applies here without modification, and the market row shows that the decay is specific rather than a general deterioration in the data.

How long a factor can be wrong

Even taking the premia at face value, the path matters, and this is where factor investing actually fails people.

Since 1993, cumulative value spread is +73.0% — a positive number over three decades. And along the way it fell 79.1 percentage points below its own running peak. Size, cumulatively +16.1% since 1993, fell 50.4 points below its peak.

No individual holds through that. The Behavioural finance subject's measured behaviour gap is largest in exactly these conditions: a strategy that underperforms for a decade is abandoned near its worst point, and the investor collects the drawdown without the recovery. A premium you cannot hold through is not a premium available to you, which is the Measuring your return subject's argument about volatility applied to a strategy rather than to an asset.

Reading an Indian factor index

India has published factor indices, and their factsheets are unusually honest documents if you read the dates.

The Nifty200 Momentum 30 selects the top 30 of the Nifty 200 on a score built from "6-month and 12-month price return, adjusted for its daily price return volatility", tilt-weighted by multiplying free-float market cap by that score, with weights capped at the lower of 5% or five times the stock's free-float weight.

Its reported returns, against the Nifty 50:

	5 years	Since inception
Nifty200 Momentum 30	9.28%	18.30%
Nifty 50	6.38%	12.12%
Difference	+2.90	+6.18

Now the dates on the same factsheet. Launch date 25 August 2020. Base date 1 April 2005.

So "since inception" is mostly a backtest. Fifteen of the twenty-one years are simulated — the index did not exist, nobody could hold it, and its rules were chosen with knowledge of that period. The 6.18-point since-inception advantage is not a track record.

The five-year column is largely live, covering September 2021 to September 2026, and it shows a 2.90-point advantage. That is the number worth looking at, and it is less than half the backtested one — which is the characteristic pattern.

Three further cautions on any such comparison.

Index returns are not fund returns. A fund tracking a factor index pays an expense ratio and incurs turnover costs that the index does not. Chapter 11 is about the second of those, and the Mutual funds subject covers the first.

The factor index is more volatile. Momentum's standard deviation is 19.44% against the Nifty's 13.82%. Chapter 9's ratio is the right comparison, and on it Momentum still wins over this window — 0.194 against 0.064.

And a factor index is a selected survivor. Exchanges launch indices that backtest well. The ones that did not are not on the website, which is the Technical analysis subject's data-snooping problem in a commercial form.

Working the problem

Value: 5.26% before 1993, 2.17% after, −1.16% over twenty years.

Explanation 1 — it was never there. The original finding was a product of searching the same dataset many ways. The Technical analysis subject sets out the mechanism: with enough rules tested on one history, some will look profitable by chance, and "any successful results may be spurious because they could be obtained just by chance with exaggerated significance levels." On this account the post-1993 numbers are the honest out-of-sample test, and it failed.

Explanation 2 — it was there and got competed away. Publication told everyone. Money flowed into value strategies, bidding up cheap stocks until the premium disappeared. On this account the finding was true when made and self-destroyed by being known.

Explanation 3 — it is there and this was a bad stretch. A premium with 12% annual volatility produces thirty-year samples that look like nothing fairly often. Value is compensation for a real risk that happened not to show up.

What would separate them.

Explanation 1 predicts the premium is absent in data the original researchers did not use — other countries, other periods, other asset classes. Evidence from markets not in the original sample is the cleanest discriminator, and it is why international replication is the standard test.

Explanation 2 predicts a structural break around publication and around the growth of value funds, with the decay tracking assets deployed rather than time.

Explanation 3 predicts the premium recovers, and that the bad stretch coincides with periods when the underlying risk story should have been painful.

Honestly: the three are hard to separate, and all three are partly true. The $t$-statistic falling from 3.37 to 1.10 is consistent with both 1 and 2. The thirty-year span is long enough to be troubling and short enough that 3 cannot be dismissed. Anyone claiming to know which is operating is claiming more than the data supports, and that includes anyone selling the fund.

What I would do with a factor fund today — four things, in order.

Treat the backtest as advertising and the live record as evidence. Find the launch date on the factsheet, and ignore everything before it. For the Indian momentum index that means looking at five years, not twenty-one.

Size it so a decade of underperformance is survivable. Not a conviction position. Value's 79-point shortfall from its own peak is what the tail of this looks like, and the question is not whether you believe the premium but whether you would still hold the fund in year eight.

Compare it against the plain index after costs. The factor fund's expense ratio and turnover are real and certain; the premium is contested. A 2.90-point gross advantage with a one-point cost disadvantage is a different proposition from the headline.

And prefer the diversified default unless you have a specific reason. Chapter 7's robust layer — that only undiversifiable risk is paid for — survives all of this. The contested layer is which risks those are. A broad low-cost index fund does not require you to resolve that question, and a factor fund does.

The point

Factors are long-short spreads on characteristics, and over the century to August 2026 the market premium earned 8.35% a year with a $t$-statistic of 4.55 while value earned 4.22% at 3.43 and size 1.99% at 1.83 — characteristics CAPM says should not be priced. But after the papers were published in 1992 and 1993 the market premium held at 9.19% while size fell to 0.48% and value to 2.17%, and over the last twenty years both are negative. Value's cumulative spread since 1993 is positive and fell 79 points below its own peak on the way, which is longer than anyone holds. Indian factor indices show the same pattern in miniature, with a 6.18-point since-inception advantage that is mostly backtest against a 2.90-point live one. Whether the premia were never real, were competed away, or are merely resting is not resolvable from this data.