How the factor numbers are computed

Sources

All factor and industry returns come from Kenneth R. French’s data library at the Tuck School of Business, Dartmouth: the 3-factor file (Mkt-RF, SMB, HML, RF), the 5-factor file (adds RMW and CMA), the momentum file, and the 49 value-weighted industry portfolios. The factor definitions are Fama and French (1993) for the three factors and Fama and French (2015) for the five; momentum is the Mom series from the same library. The volatility split uses Cboe VIX daily closes, averaged to a month, over the 437 months from Jan 1990 to May 2026 where the two datasets overlap. No value on these pages is interpolated, imputed or forecast.

Units: percent, not decimals

French publishes returns in percent: a value of 0.53 means 0.53%, not 53%. Every compound on these pages therefore multiplies (1 + r/100). Getting this wrong scales every number by 100 and nothing fails, so the build asserts it: if the market factor’s long-run monthly mean, currently 0.6966 over 1,199 months, ever leaves the 0.4 to 1.0 band that an equity risk premium quoted in percent per month must occupy, the page throws instead of rendering.

Missing-value sentinels

French codes a missing observation as -99.99 or -999, not as a blank. Averaging one of those into a return series would move a mean by tens of percentage points and would leave no trace. In this vintage of the parquet the sentinels have already been converted to SQL NULL, and the build asserts it: every value at or below -99.98 in any of the four monthly files raises an error. For scale, the most negative return in the 49 industry portfolios is -62.1%, Business Supplies in Jun 1930.

Compounding

Cumulative returns are geometric: the product of (1 + r/100) across months. The compound annual rate is that product raised to 12/n, minus 1. An arithmetic mean of monthly returns, annualised by multiplying by 12, overstates the compound outcome of any volatile series, and the more volatile the series the larger the overstatement. Both quantities appear in the tables, labelled: the mean and its t-statistic are arithmetic, because that is what the factor literature tests; the growth of $1 and the CAGR are geometric, because that is what an investor would have experienced.

Two consequences worth stating plainly. First, compounding Mkt-RF gives the growth of the excess-return series itself, which is not the same number as the market’s compound total return minus the compound return on cash. Second, compounding a long-short spread as though $1 were invested in it is a convention, not a portfolio: the spread is zero-cost, and the level of a long-short index has no dollar interpretation.

Two Mkt-RF series that do not agree

French publishes the 3-factor and 5-factor files separately, and they disagree. Over their 755 overlapping months, Mkt-RF differs in 167 of them, by at most 0.08 percentage points. SMB differs in 742 months, by up to 3.52 points: French builds the 5-factor SMB as the average of the small-minus-big returns from three separate 2x3 sorts, and the 3-factor SMB from one. HML differs in 0 months and RF in 0. This module takes Mkt-RF, RF, SMB and HML from the 3-factor file, which is the long series back to Jul 1926, and RMW and CMA from the 5-factor file, which begins in Jul 1963. The SMB figures here will therefore not reconcile against the 5-factor file.

Windows are derived, not fixed

“The last 20 years” is the trailing 240 months ending at the last month present in the data (May 2026 in this build), computed at build time. The era split is the month the HML cumulative index reached its all-time high, Dec 2006, also computed from the data: it is not a date chosen to flatter the result. Every rolling ten-year figure uses complete 120-month windows only; a partial window would annualise a short stretch and read as a real ten-year number.

The 49 industry portfolios

40 of the 49 industries carry all 1,199 months of the panel. 9 begin after Jul 1926: Personal Services (Jul 1927), Business Supplies (Jul 1929), Rubber and Plastic Products (Jul 1930), Defense (Jul 1963), Candy and Soda (Jul 1963), Precious Metals (Jul 1963), Fabricated Products (Jul 1963), Computer Software (Jul 1965), Healthcare (Jul 1969). Compounding runs over observed months only, and each industry is benchmarked against the market compounded over exactly the same months, so a portfolio that starts in the 1960s is never scored against a market window it did not live through.

Rubber and Plastic Products is the only industry with an interior gap: 12 months inside its span carry no observation. Those months are dropped, so its cumulative index chains across the gap. 2 of the 49 carry a month above +100%, 4 such months in all: Business Supplies in 3 of them, the largest +300.0% in Aug 1932, Coal in 1 of them, the largest +125.4% in Aug 1932. Those are real values in French’s depression-era data, not sentinels, and they are left in. Business Supplies carries the highest annualised volatility of the 49 at 52.1%.

The industry names are French’s own (the Siccodes49 definitions) and the parquet columns are his abbreviations, which do not always read as the definition does: Hardw is Computers, Chips is Electronic Equipment, Paper is Business Supplies, Fin is Trading. Firms are assigned by SIC code. The parquet carries the 49 return series and nothing else, so which firms sat inside a portfolio in a given month cannot be read off this data.

What this cannot support

• No transaction costs, spreads, shorting fees, borrow constraints, capacity limits or taxes are in any of these numbers. The long-short factors are paper portfolios. Their returns are an upper bound on what a real investor could have earned, and this data cannot say by how much.
• Every return series here comes from French’s four US files. Nothing here speaks to Japan, Europe or emerging markets.
• Fama and French selected and published these factors from samples inside these 1,199 months (1993 for the three factors, 2015 for the five). The t-statistic of 3.43 on HML’s full-sample monthly mean is not an out-of-sample test of anything; it is a description of the series the literature was built by searching.
• The page measures when the factors stopped paying. It cannot say why. That HML’s index has been below its Dec 2006 peak for 233 months is a fact; whether the premium was arbitraged away, whether book value stopped measuring what it measured, or whether this is a long unlucky run in a series whose full-sample annualised volatility is 12.3%, this data cannot distinguish.
• The VIX split is a split of realised returns into two halves of the sample. It is not a conditional model, it is not a trading rule, and it supports no causal claim.

Coverage as built

The factor rows are excess or long-short returns; the industry row is the equal-weighted mean of the 49 portfolios’ own mean monthly total returns, which are not net of cash. The two are in the same unit but they are not the same quantity, and the column should not be read down.