0005 — The public factor files cannot sign the profitability or investment premium

Status: scoped instrument finding. The French-file designs, panels, windows, and estimands tested here lacked resolution for decision-material RMW and CMA effects. “Closed” does not reach alternative data, conditional models, different estimands, longer windows, or lower-variance designs. See decision 0010.

Date: 2026-08-12. Status: accepted. Rescoped by decision 0010 and by decision 0009: the measured floors below are unchanged, but “closed, not paused” and “provably will not” overstated them. The honest claim is about an instrument — no re-pooling, re-weighting, re-windowing or re-blocking of these files can sign a premium below about 2.6 pp/yr at 80% power — and it does not forbid a different estimand, a conditional or panel design, a non-French construction, or a lower-variance estimator.

Context

Experiment 001 left HML, UMD and RMW unresolved, not because their premia were shown to be absent but because 16 of its 20 factor × era cells held a premium smaller than their own window could detect at 80% power. More United States history cannot fix that: the US post-publication window is the length it is and it ends at the frozen sample boundary.

Experiment 005 was designed to settle it in one of two directions, with both branches of its falsifier frozen before any number existed. It re-ran Experiment 001’s post-publication grid for HML, RMW and CMA across the US, developed-ex-US and emerging Ken French files over the same frozen eras, pooled them under a cross-region joint block bootstrap that preserves contemporaneous cross-region correlation, and — this is the point — measured the effective sample size that pooling actually bought rather than assuming it.

The measurement, on the full post-publication era of each factor:

Branch (b) of the frozen falsifier fired for RMW and CMA: the measured pooled minimum detectable effect at 80% power is above this repository’s 2.0 pp/yr materiality threshold. It fires at the point estimate, across the whole 90% sampling interval of the MDE, and across the whole Phase 1 systematic volatility band. RMW’s pooled premium of +2.53 pp/yr is smaller than the smallest premium its own pooled window can resolve, and RMW is the factor pooling helped most, because its three regions are the least correlated in the grid.

Branch (a) fired for HML, which is now exploratory. UMD was left out of Experiment 005’s frozen universe, on the stated ground that no regional momentum file existed here — a fact about the repository, not the data. Experiment 006 acquired the three files Ken French does publish and ran the identical design on UMD: branch (a) fired there too, on a pooled +7.33 pp/yr against a measured pooled detection threshold of 4.98 pp/yr and 1.33 effective regions out of three. That threshold is now the worst pooled figure in this repository, and it raises the floor below rather than lowering it.

Decision

The Ken French public files cannot sign the RMW or CMA premium, and no rearrangement of them will. Both are rejected against the predeclared falsifier on these series over these windows under this construction. This is a statement about an instrument, and it closes one route rather than the question — clause 8 of decision 0010.

The general statement the measurement supports:

On the Ken French public factor files, across every independent region the library distributes, a post-publication premium between zero and about 2.6 percentage points a year cannot be signed at 80% power, however the regions are pooled.

That floor — 2.62 pp/yr, RMW’s pooled figure, the best across all nine of Experiment 005’s pooled cells and all twelve once Experiment 006’s three are added — is above the 2.0 pp/yr materiality threshold this repository uses. It is a property of the available data, not of the factors. Momentum’s own pooled floor is 4.98 pp/yr, roughly twice as high, because its three regions are the most correlated of the four factors and its volatility the largest.

What rejected means here, precisely. It is the predeclared falsifier firing on these series over these windows under this construction. It is not a claim that the profitability or investment premium is zero, and it must never be reported as one. The honest statement is that the publicly available evidence cannot sign the premium either way, and that more of the same evidence will not either.

What the measurement reaches.

• Rearranging these files is answered, not forbidden. Re-pooling, re-weighting, re-windowing and re-blocking were all tried: the floor holds under the frozen 12-month block with 6- and 24-month neighbours and the Politis–White automatic length reported for every cell, under equal and inverse-variance weights, and with and without the US leg. None changed a verdict. Offering one of those as a new answer is repeating a measured result, which is why it is pointless rather than why it is prohibited.
• A product may not be promoted on RMW or CMA premium grounds, because the chain a shareholder receives is premium × delivered loading − cost and an unsigned premium makes the product’s own quality irrelevant to the decision (decision 0004). This is a promotion gate. It does not stop a sleeve, tilt or product from being registered, measured, costed or priced (decision 0010, clause 5).

What the measurement does not reach, and therefore does not close: a different estimand; a conditional or panel design; a non-French construction on a different universe and a different vendor’s accounting data; and a lower-variance estimator on these same files. The framework’s own literature ledger records a published replication rate moving 21.1 points on a change of estimand with no new data, so this is not a theoretical gap. Each needs a new frozen specification; none needs this record superseded.

What would reopen it

Each is a measurable condition, not a hope. Any one of them reopens the specific factor it names, under a new frozen specification.

What does not reopen it: a new Ken French vintage of the same files, a different block length, a different pooling weight, a different era boundary inside the same sample, the 2026-01-onward holdout (six to eight months against a 2.6 pp/yr floor), or the 2013–14 CRSP vintage that would settle the Phase 1 band — that band is checked cell by cell in both experiments and changes no conclusion anywhere.

Alternatives considered

Leave RMW and CMA unresolved again. Rejected, and this is the decision’s whole point. unresolved absorbs effort indefinitely: it reads as “come back with more data”, and Experiment 005 was built specifically to determine whether more of the available data exists. It does not. A measured statement that the question cannot be answered from public data is more useful than a third invitation to try.

Report RMW as exploratory on the grounds that its pooled interval [+1.07, +3.96] excludes zero and its premium clears 2.0 pp/yr. Rejected, and the rejection rule was frozen in advance precisely so this could not be reported as a finding. Its premium is below its own detection threshold, which means the interval excluding zero is not evidence the window can carry; 62% of its US premium is the single year 2021; and dropping the pooled composite’s best calendar year takes it to +1.79, below materiality. Clause (a5) failed on its own terms before branch (b) was reached.

Close HML too, for consistency. Rejected. Branch (a) fired for HML on every one of its five clauses, and the falsifier’s whole design was that both branches be decisive. Closing a factor whose premium exceeds its own measured detection threshold would be applying the rule to the conclusion rather than the evidence.

Extend the closure to UMD. Rejected at the time as unsupported, because UMD had never been tested regionally, and overtaken by events: the data acquisition this record named as the missing step was made, and Experiment 006 signed UMD’s premium under branch (a). UMD is exploratory, not unresolved and not closed. Its remaining obstacles are turnover, cost and a crash its three regions share — none of which this record’s closure rule addresses.

Consequences

• Experiment 001’s prioritisation of RMW as “the one worth looking at first”, on the grounds that it retained 96% of its premium, is superseded. It retained its premium, and its premium is still smaller than the smallest one three regions of public data can resolve.
• The design map’s Map C promotion conditions for profitability and investment are replaced by the reopening conditions above.
• The programme’s remaining live question for factors is no longer the premium. It was recorded here as the long-only capture fraction; that is superseded. Experiment 007 measured it and then showed it to be an HML loading rather than a multiplier, so the edge decomposition’s 0.40 is deleted rather than replaced. What is live for HML is now a product on a prior-window comparator, and whether a short-window loading forecasts anything.
• Decision 0004 stands: no sleeve is promoted. A factor reaching exploratory is not a sleeve, and every value product still has to pass Experiment 002’s frozen promotion protocol on its own terms.
• All figures behind this record are gross of transaction costs, shorting costs, borrow, fees and taxes, on academic zero-investment long-short research portfolios a retail investor cannot implement. They are upper bounds of unknown tightness, and the pooled ones are looser than the US ones because emerging-market shorting is harder and dearer than US shorting.