06-reference/research

dsr within cluster variance primary verification

2026-10-30·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
deflated-sharpeeffective-nmultiple-testingonc-clusteringprimary-source-verification

Primary-confirmed: the variance is over the K cluster-representative Sharpes, and the alternative reading is the PERMISSIVE one, not the harsher one

The question

"Verify the within-cluster aggregation rule for V[SR_n] against the López de Prado & Lewis (2019) primary: is the variance taken over cluster-representative Sharpes from aggregated cluster return series, or over all N_literal trial Sharpes while N uses N_eff?"

This is open-follow-up #1 of [[2026-08-31-dsr-effective-n-estimator-precommit]], whose step 5 asserted the first reading from secondary sources after two PDF extractions of the primary failed. The backlog entry's stated stake was that the alternative reading "gives a generally larger V[SR_n] and therefore a harsher hurdle." That premise turns out to be backwards, which is the second finding here.

What we already know (from the vault)

What the web says

All quotations below are from my own pdftotext -layout extraction of the November 1, 2018 SSRN working paper, fetched this run via curl + UA spoof to the smallake.kr mirror (25 pages, extraction clean). Page references are to that PDF's own printed page numbers.

Convergences and contradictions

Synthesis for RDCO

The pre-registration does not need amending on the question asked. It needs three mechanical additions the primary supplies and step 5 is silent on, and one correction to a stated rationale. Step 5 stands verbatim; promote its evidence tier and stop re-queueing it. What the primary adds is implementation detail that changes the number: the weighting scheme, the frequency normalization, and the behaviour at small or singleton clusters. The weighting scheme is not a free choice — the paper names minimum-variance and then, in Appendix A.4, tells you to approximate it with inverse-variance weights (optionally Sherman-Morrison-corrected) precisely because within-cluster covariance is ill-conditioned by construction. In my runs the two choices moved SR₀ by −6.5% to +1.7% depending on cluster shape, which is the same order as a whole extra trial in N_eff. Leave it unpinned and the implementer picks the hurdle. Pin it: use Appendix A.4's getIVP with the extended Sherman-Morrison term, fall back to plain inverse-variance when the extended term produces a non-positive weight sum, and never a raw pinv(Σ_k) solve, which is the convenient route and the unstable one. The frequency normalization is a no-op only under a condition RDCO has already pre-registered and should now state as the reason: step 1's pinned units (per-return-observation Sharpe, same bar frequency, same cost model across sweep and confirm) make Frequency_k = Frequency_k* for every cluster, so E[V[{ŜR_k}]] = V[{ŜR_k}] directly. If the sweep ever spans mixed bar frequencies or mixed trial spans — which the reseal machinery in [[2026-10-29-dsr-cross-block-accounting-after-reseal]] makes a live possibility, since R3 re-evaluates trials on a pinned window of growing length — the division by Frequency_k* becomes load-bearing and its omission silently rescales the hurdle. Add the per-cluster T_k, Years_k, Frequency_k and Frequency_k* to the ledger's logged fields so the no-op is visibly a no-op rather than an assumption.

The small-cluster behaviour is where step 5 is actually underspecified, and it bites in both directions. ONC searches k = 2, …, N−1, so K = 1 is not in its output space: the pre-registration's clamp(..., 1, N_literal) has an unreachable lower bound, which is harmless, but the paired assumption that a maximally-collapsed ledger yields N_eff = 1 is wrong — it yields 2. More consequentially, R3 of the cross-block rule already creates singleton clusters by fiat (a trial that cannot be re-evaluated on the reference window "counts as its own cluster"), and the primary says singletons are hard for ONC to recover and excludes them from its own validation design. A singleton cluster's "aggregate" series is just that trial's own return series with weight 1, which is well-defined and should be stated so, but it also means every singleton injects an un-smoothed, full-noise Sharpe into a variance computed over very few points. That compounds the estimator-noise problem below, and it is the mechanism by which R3's conservative-looking exclusion rule can move V[{ŜR_k}] in either direction. Pre-register the handling explicitly rather than letting it fall out of the code.

The finding neither parent brief raises, and the one I would act on first: V[{ŜR_k}] is a variance estimated from K points, and at the K values RDCO will actually see, it is noise. Across my runs the run-to-run coefficient of variation of √V[{ŜR_k}] at K = 10 was 0.21 to 0.26. Since SR₀ is linear in √V, the hurdle inherits a ±21-26% one-sigma wobble from sampling alone, with nothing changing in the strategy space. Set against the ratchet logic in [[2026-10-29-dsr-cross-block-accounting-after-reseal]] — where R4 fires on a 0.14% dispersion fall at N_eff = 100 — a 25% sampling CV means the SR₀_gate = max(SR₀_new, SR₀_prev) ratchet will be driven almost entirely by estimator noise rather than by laundering, and after a few reseals it will have ratcheted up to roughly the running maximum of a noisy draw. That is conservative, so it fails safe, but it is conservative for the wrong reason and it will be mistaken for signal. Two things follow. First, report an interval, not a point: bootstrap V[{ŜR_k}] over clusters and log the resulting SR₀ band next to the point estimate, so a marginal DSR is visibly marginal. Second, this is the strongest argument yet for the dimension-reduction route the cross-block brief identified as the way to "earn" the discount — a ledger that clusters to K = 40 has a √V CV near 11% rather than 25%, and K ≥ 30 would make the pre-registration's own ≥30-trial fallback floor coherent with the quantity it is protecting. Note that under R9 (M_max = 45 at the current ~1,005-bar window) the clustering refuses on realistic v0 sweeps anyway, in which case N_eff = N_literal and this variance question is dormant — but the moment a dimension-reduced ledger turns the discount on, the dispersion estimate becomes the noisiest input in the gate.

One process finding, since the dispatch itself is evidence. This question was auto-promoted from a backlog entry on 2026-09-02 and dispatched 2026-10-30, having been answered on 2026-10-29 by a brief that said so in its own "why this is in the vault" section. The duplicate work was not wasted — it produced an independent extraction, corrected the backlog's directional premise, corrected the sibling's stated mechanism, and surfaced the estimator-noise and weighting-scheme issues — but that is luck, not design. The retirement step the queue is missing is specific: when a brief's "why this is in the vault" claims closure of another brief's numbered open follow-up, the corresponding backlog row should be closed in the same write. That is a one-line check, and the /deep-research backlog is already known to self-inflate.

Why this is in the vault

This closes open-follow-up #1 of [[2026-08-31-dsr-effective-n-estimator-precommit]] at primary tier and hands validation.py's unbuilt deflated_sharpe(...) three specific build inputs it was missing: the Appendix A.4 inverse-variance weighting (not pinv(Σ_k)), the per-cluster frequency normalization with its no-op condition logged, and explicit singleton-cluster handling. It also corrects the direction stated in that brief's own follow-up text, which claimed the unverified alternative was the harsher reading when it is the permissive one — the opposite of the risk the follow-up was written to guard.

Open follow-ups

Related

Sources

Vault:

Web / primary:

Own computation (reproducible, synthetic data only):