Fama–French factor reproduction: the Phase 1 ingestion gate

Question. Does this repository’s data path — download, raw-byte cache, zip extraction, multi-table parsing, percent-to-decimal conversion, sample-boundary selection and monthly statistics — reproduce a precisely identified published table?

Decision it informs. Whether any downstream experiment may draw strategy conclusions from Ken French data. Out of scope: whether any factor is real, investable, or worth holding.

---

Conclusion

The gate is unresolved against its predeclared tolerance. Thirteen of fifteen gating cells reproduce; the two that do not are the standard deviations of HML and RMW.

• All five monthly means reproduce, the largest miss being RMW at +0.0142 pp/month against a declared tolerance of 0.02.
• All five t-statistics reproduce, and all ten cross-factor correlations.
• HML’s standard deviation is 3.0% too low (2.7926 vs a printed 2.88) and RMW’s is 5.1% too high (2.2490 vs 2.14), against a declared tolerance of 0.05 pp/month.

Exact reproduction is not available at any tolerance: Ken French publishes no vintage archive, so the 2013–14 CRSP vintage the paper used cannot be obtained. The tolerance was declared before any number was computed and has not been touched.

Consequence. The ingestion path is trustworthy for first moments and for cross-sectional structure. It is not established to reproduce second moments of HML and RMW to better than about 5%. Any downstream result that divides by an HML or RMW volatility — a Sharpe ratio, a volatility-scaled sleeve, a risk-parity weight, a covariance matrix, a Kelly fraction — inherits a systematic ±3–5% uncertainty that is not sampling error and will not shrink with more data. Report those quantities with the band attached, or state that it was not propagated.

---

What was targeted, and where the published values came from

The commissioning brief named “Fama and French (2015) Table 1”. Table 1 of that paper is the average excess returns of the 25 test portfolios, not the factors. The table holding the mean, standard deviation and t-statistic of the five monthly factor returns is Table 4, Panel A, first block (2 × 3 Factors), over July 1963 – December 2013, 606 months, in percent per month. The brief’s description matched Table 4 exactly; only the number was wrong, and Table 4 is what the frozen specification targets.

The typeset Elsevier PDF was retrieved (sha256 5712e97e…, 2026-08-12) and carries the running head, submission dates and copyright line, so it is the published article rather than a preprint. Eight other routes — ScienceDirect, SSRN, Chicago Booth, MIT DSpace, CiteSeerX and three mirrors — returned 403, 404 or 429.

Cross-checked, not merely extracted. The paper’s own prose in Section 4 independently restates nine of the table’s values and every one agrees with the extracted table. A Wayback snapshot of the November 2013 working paper carries the same table over 594 months and is used only as a second, non-gating reference point.

---

Results

From the pinned file over 1963-07…2013-12, 606 months, percent per month. Run f8972356df0648cf8867b1c742261bda, specification hash 5bf03adea820….

Tolerances declared before anything was computed: mean 0.02, standard deviation 0.05, t-statistic 0.30.

Which series are already excess, and which are not. Mkt-RF, SMB, HML, RMW and CMA are all already excess or long-short returns, and RF was not subtracted from any of them. Subtracting it a second time is the classic error here and it is invisible in the output, because the result is still a plausible small monthly number. The experiment measures it rather than asserting its absence: doing so would shift every factor mean by −0.4153 pp/month (−4.98 pp/yr) and flip every long-short sign.

Annualisation conventions, stated because they are silent traps: premium = 12 × monthly mean (arithmetic, not a compound growth rate); volatility = sqrt(12) × monthly SD. Both assume serially independent monthly returns, which these are not — the gap between the iid and HAC standard errors is 7% to 18%, which is the size of that assumption. The published t-statistics are the iid ones, so those are what the gate compares.

---

Diagnosis of the two failing cells

Verified.

1. The failure is confined to second moments of two factors. Variance ratios (ours ÷ printed): Mkt-RF 0.997, SMB 0.984, CMA 0.989, HML 0.940, RMW 1.104.
2. The same two cells fail against both printed vintages, by the same amount — HML −0.0851 and RMW +0.1136 against the 594-month draft, versus −0.0874 and +0.1090 against the published table. Two independently typeset vintages a year apart give the same discrepancy, so this is not the 2013→2014 revision cycle.
3. The correlation structure is unchanged. All ten correlations reproduce within 0.06 and nine within 0.033, against the paper’s own standard error of 0.04.
4. HML’s and RMW’s variance is concentrated where restatement bites hardest. Of RMW’s 606-month variance, 49.5% comes from the 36 months of 2000–2002; HML 22.0%, against Mkt-RF’s 10.3%. RMW’s 2000s decade standard deviation is 3.90 against 1.48–1.69 in every other decade.
5. The sample boundary and file integrity are not the cause. The window holds exactly 606 strictly increasing, gapless, duplicate-free months with no missing values and no sentinel leakage. Shifting it one month forward moves the Mkt-RF mean to 0.4962; one month back selects 605 rows and is caught.
6. Rounding cannot explain it. Recomputing the published t-statistics from the published rounded inputs gives 2.74 / 2.33 / 3.16 / 2.88 / 4.04 against printed 2.74 / 2.31 / 3.20 / 2.92 / 4.07, confirming the printed rows are independently rounded from unrounded inputs. Printing precision accounts for ±0.005, not 0.087 or 0.109.

Inference, not a finding. CRSP and Compustat restate delistings, share counts, prices and accounting data retroactively. HML and RMW are built from the extremes of book-to-market and operating profitability — precisely the cohort of small, unprofitable, high-growth firms around the dot-com period where twelve years of restatement change the most. That would perturb the extreme portfolios’ monthly returns without much changing their long-run average, which is the pattern observed. Distinguishing it from a subtle construction difference requires the 2013–14 vintage, which does not exist publicly. The Ken French page records no definitional change, but a page that does not mention a change is not evidence that none occurred.

---

What passing this gate does NOT establish

Stated because the failure mode of a green integration test is that it gets quoted as evidence of something it never tested.

• It does not verify portfolio accounting, rebalancing, transaction costs, tax lots, corporate actions, insolvency behaviour, missing-data policy, look-ahead protection, or optimisation constraints. None of them is exercised.
• The published factor file already contains the authors’ own calculated returns. Matching their table shows we read and summarised their numbers correctly; it does not show we could have computed them.
• It can match despite compensating errors — two offsetting mistakes in the unit transform and the window would produce the same table.
• Every figure is gross, because the source series are. The long-short series are not investable.
• A sha256 proves which file was used, never what was available at an earlier date. Nothing here is point-in-time.

---

Where the band changes a conclusion: nowhere, and that is structural

The band is carried as a separate systematic band and never combined with a sampling interval. Every downstream primary metric and falsifier clause is a function of the mean, which reproduced for all five factors; the band moves only volatility, Sharpe and the minimum detectable effect. Checked cell by cell in factor persistence, every reading holds at both ends of the band — including the pooled MDE that decides branch (b), where it is roughly ±1.5% because two of three legs are unaffected.

Five series carry no measured band at all, which is weaker than a band of zero: the two regional five-factor files, and all three momentum files, never gated against any printed table in any region.

---

Open questions

• Can any 2013–14 vintage of the file be obtained — a co-author’s archive, a replication package, an institutional mirror? That is the single observation that would settle the diagnosis, and it changes no conclusion anywhere.
• Do the differences persist in the 2 × 2 and 2 × 2 × 2 × 2 blocks, which use different portfolio intersections? The library distributes only the 2 × 3 file.
• Is the discrepancy localised to 2000–2002 or spread? Answering it requires the published monthly series, which was never distributed.

Reproduce it

cd research
uv run python -m portfolio_edge.experiments.exp_phase1_ff_reproduction --view-results
uv run pytest tests/integration/test_phase1_reproduction.py      # offline
uv run pytest -m network tests/integration/test_phase1_reproduction.py

The source vintage is F-F_Research_Data_5_Factors_2x3_CSV.zip, sha256 cbc37248…, Last-Modified 2026-08-03, retrieved 2026-08-12, whose own preamble reads “This file was created using the 202606 CRSP database.” The specification refuses to run against a file whose sha256 is not the pinned one: when Ken French publishes a new vintage this experiment aborts rather than reporting numbers, and a new specification must be frozen against the new hash — which is the point.