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)
- The pre-registration already specifies the cluster-representative reading, and flagged it as unverified. [[2026-08-31-dsr-effective-n-estimator-precommit]] step 5:
V[SR_n]= cross-sectional variance of the cluster-representative Sharpes, "one per ONC cluster, the cluster's aggregated return series," consistent with its ownN_effclustering. Its open-follow-up #1 names the alternative and says the mechanic is asserted from secondary sources only. - A sibling brief reached the primary yesterday and closed this follow-up — one day before this dispatch was issued. [[2026-10-29-dsr-cross-block-accounting-after-reseal]] reports a successful deep-read of the same paper via the smallake.kr PDF mirror, quotes §8.2's minimum-variance weights and frequency normalization, and pre-registers R7: "
V[{SR_k}]stays on cluster representatives... Log the all-trials dispersion alongside it for audit, flagged not-for-use." Its line 103 states the closure explicitly. This brief is therefore a second, independent extraction of the same primary rather than a first reach — which is the right posture for a claim that was going to be coded either way, but the backlog entry should have been retired, and the duplication is itself a process finding. - The sibling brief's stated mechanism for the direction is the one place it is loose. It argues from equicorrelation: for M equicorrelated trial Sharpes
E[S²] = σ²(1−ρ), so the all-trials dispersion shrinksSR₀by√(1−ρ)— quoted as 0.707 at ρ = 0.5. The conclusion is right; the mechanism is not the dominant one, because in a block-correlated ledger the average pairwise correlation is small even when intra-cluster correlation is near 1. Quantified below. - The adjacent primary-verification work is already done and consistent. [[2026-09-04-dsr-denominator-sr0-vs-srobs-resolution]] confirmed from the DSR primary that the variable list reads "V[{ŜR_n}] is the variance across the trials' estimated SR and N is the number of independent trials" — the same index coupling this brief finds in the Lewis paper. [[2026-09-05-afml-embargo-cscv-primary-source-verification]] established the vault's house style for tiering a primary read against a mirror.
- The guardrails that make this question matter are already pre-registered. Both the
N_effguardrails ("take the maximum, clamp toN_literal, refuse below ~30 trials") and the laterSR₀_gate = max(SR₀_new, SR₀_prev)ratchet exist precisely because every knob in this formula points toward a lower hurdle.
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.
- §8.2's title and first sentences settle it outright. The section is headed "ESTIMATION OF THE VARIANCE OF CLUSTERED TRIALS, E[V[{ŜR_k}]]", and reads: "Upon completion of the clustering above, ONC has successfully partitioned our N strategies into K groups... In this section, our goal is to utilize the clustering to reduce the N strategies to K ≪ N cluster-level strategies. Upon creation of these 'cluster strategies,' we derive our estimate E[V[{ŜR_k}]] for each k = 1, …, K" (p.11). There is no variant in which the variance is taken over the N literal trials.
- The aggregation rule is explicit, and so is the weighting scheme. "For a given cluster k, the goal is to form an aggregate cluster returns time series S_k,t. This necessitates choosing a weighting scheme for the aggregation. We choose the minimum variance allocation... to mitigate the adverse effects of any strategies with larger variance," giving
w_k = Σ_k⁻¹1 / (1'Σ_k⁻¹1)andS_k,t = Σ_{i∈C_k} w_k,i·r_i,t(p.11). Each cluster's Sharpe is then estimated fromS_k,t, annualized at that cluster's ownFrequency_k = T_k/Years_k, andE[V[{ŜR_k}]] = V[{aŜR_k}] / Frequency_k*whereFrequency_k*is the frequency of the selected strategy — "Otherwise, SR* would not be estimated on the same frequency as ŜR" (pp.11-12). - The alternative reading is not just unsupported, it is notationally incoherent with the paper. §7's own definition runs the variance and the count over the same index set: "SR* = √V[{ŜR_k}]((1−γ)Z⁻¹[1−1/K] + γZ⁻¹[1−1/(Ke)]) where V[{ŜR_k}] is the variance across the trials' estimated SR, K is the number of independent trials... and k = 1, …, K" (p.8). The False Strategy Theorem it rests on is stated "Given a sample of IID-Gaussian Sharpe ratios, {ŜR_k}, k = 1, …, K" (p.7-8) — so the variance is over the set the theorem's independence assumption applies to. Mixing an N-indexed variance with a K-indexed count would use two different populations in one equation.
- The paper is explicit that K ≤ N and that this is the whole point. "The number of independent trials K is less or equal to the number of trials N" (p.9), and E[K] / E[V[{ŜR_k}]] are the two "meta-research variables... intrinsic to the research process itself" that "cannot be estimated from the selected strategy" (pp.8-9). The conclusions restate it: "This requires the estimation of two meta-research variables... We estimate these two meta-research variables with the help of the ONC algorithm" (p.13).
- Appendix A.4 prescribes a specific robust substitute for the matrix inverse, which the pre-registration does not mention. "Note that the intra-cluster correlations will be high by design, and thus Σ_k may be ill-conditioned and difficult to invert. In practice, one could choose to approximate the weights by setting w_k,i proportional to 1/σ_i²... If more accuracy is desired," use the Sherman-Morrison rank-one update under an average off-diagonal correlation ρ, giving
w_k,i ~ 1/σ_i² − ρ·Σ_{j∈C_k}(1/σ_j) / ((1+(N_k−1)ρ)σ_i)(pp.17-18, Snippet 4). This is an inverse-variance allocation, not a full min-variance solve. - ONC's search range makes K = 1 unreachable, and singletons are a known weak spot. The base algorithm loops "k = 2, …, N−1" (p.10), and the Monte Carlo design notes "each block must be of size M ≥ 2, as blocks of size 1 are difficult to identify as a cluster" (p.12).
- The published version is still unread. Quantitative Finance 19(9), 1555-1565, DOI 10.1080/14697688.2019.1622311 is paywalled on tandfonline; RePEc confirms the pagination and the subscriber restriction. Not retried, per the paywall rule. The author's own site posts a later lecture, "Detection of False Investment Strategies through FWER and FDR" (SSRN 3799803, linked from quantresearch.org/Lectures.htm); SSRN's direct-download endpoint returned a Cloudflare interstitial rather than a PDF, so it is unread too and left as a route, not a result.
Convergences and contradictions
- Convergence, now at primary tier: step 5 of the pre-registration is correct as written. The vault's secondary-sourced guess, the sibling brief's independent read, and my own extraction agree. The claim can be promoted from "asserted from secondary sources" to "primary-confirmed against the Nov 2018 working paper, published version unverified."
- Contradiction with the backlog entry's own rationale, and it is the operationally important one. The entry says the alternative gives a "generally larger V[SR_n] and therefore a harsher hurdle," which would make the open follow-up a conservatism check. It is the reverse. In a block-correlated Monte Carlo (T = 1,000 bars, trials as
√ρ·f_k + √(1−ρ)·ε, 1% bar vol, cluster count in theN_effslot in both arms), the cluster-representative variance was larger than the all-trials variance in every configuration tested, andSR₀correspondingly harsher: +4.8% at 10 clusters × 20 trials with ρ_in = 0.7, +14.1% at 20 × 100 with ρ_in = 0.95, +10.9% with one of ten clusters carrying real edge, +26.6% at 10 balanced clusters of 168 with ρ_in = 0.9, and +35.3% to +52.8% when grid density is concentrated in one junk cluster (400 cells in one cluster, 20 in each of nine). The alternative reading is the permissive one, so the pre-registration was already erring in the right direction and the risk of having guessed wrong was a false-positive risk, not an excess-conservatism one. - Contradiction with the sibling brief's stated mechanism, same conclusion. Equicorrelation shrinkage cannot account for most of that gap. With 10 blocks of 168 at ρ_in = 0.9, the average pairwise correlation across all N is only ρ̄ = 0.090, so
√(1−ρ̄)predicts just +4.8% — against +26.6% measured. The dominant channel is different and larger: minimum-variance aggregation of many highly-correlated trials cuts the cluster series' volatility by more than its mean, so|ŜR_k|runs above a typical member's|ŜR_i|(0.0274 vs 0.0247, +11%, in the 20 × 100 ρ_in = 0.95 arm) and the cross-sectional dispersion scales up with it. Second channel: one-vote-per-cluster removes the grid-density weighting that lets a densely-sampled junk region dominate the all-trials dispersion — which is why the unbalanced arms show the largest gap. Both are properties of the search design, not of the strategy space, which is the argument for the paper's rule beyond deference to it.
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
- Does the published Quantitative Finance 19(9) 2019 §8.2 differ from the Nov 2018 working paper? Both the sibling brief and this one read only the preprint. The remaining untried routes are the author's own 2021 FWER/FDR lecture (SSRN 3799803, Cloudflare-blocked on direct download, reachable via Playwright navigate per the vault's own workaround) and the "Machine Learning for Asset Managers" (2020) chapter on testing-set overfitting, which restates ONC in book form.
- What is the sampling distribution of
V[{ŜR_k}]as a function of K, and at what K does the hurdle band narrow enough to gate on? My ~21-26% CV at K = 10 is a Monte Carlo result on synthetic block-correlated trials, not a derivation, and the primary offers no standard error for its own estimator. A derived or tabulated answer would set a minimum-K admission bar for the discount, replacing the pre-registration's unmotivated ≥30-trial floor. - Does minimum-variance within-cluster aggregation bias
V[{ŜR_k}]upward by construction, and is that bias the paper's intent or an artifact? The volatility reduction that raises|ŜR_k|above member Sharpes is a mechanical consequence of the weighting choice. If upward, the rule is conservative by accident and the magnitude should be characterized before it is relied on; the paper states only a robustness motive ("mitigate the adverse effects of any strategies with larger variance") and never addresses the effect on the dispersion it is feeding. - How should singleton and tiny clusters enter
V[{ŜR_k}]? R3 of the cross-block rule manufactures singletons, the primary excludes blocks of size 1 from its validation design, and a singleton contributes an unsmoothed Sharpe to a K-point variance. A winsorizing or minimum-cluster-size rule may be needed, and it changes the hurdle. - Does the Extreme Value Theory derivation survive substituting K for an IID count when the cluster representatives are themselves only approximately uncorrelated? Carried unresolved from [[2026-10-29-dsr-cross-block-accounting-after-reseal]]; this brief adds that the cluster representatives inherit residual inter-cluster correlation from ONC's "low inter-cluster correlation" objective, which is a bound, not a zero.
Related
- [[2026-08-31-dsr-effective-n-estimator-precommit]] — the pre-registration under test; its step 5 is confirmed here at primary tier, its follow-up #1's stated direction is corrected, and three build inputs are added
- [[2026-10-29-dsr-cross-block-accounting-after-reseal]] — the sibling brief that reached the same primary one day earlier; its R7 is confirmed, its equicorrelation mechanism is corrected, and its R4 ratchet is shown to be dominated by estimator noise
- [[2026-09-04-dsr-denominator-sr0-vs-srobs-resolution]] — the companion primary verification on the DSR denominator; same index-coupling reading of the variable list
- [[2026-09-05-afml-embargo-cscv-primary-source-verification]] — the vault's precedent for tiering a mirror-sourced primary read against a paywalled published version
- [[2026-06-01-cpcv-deflated-sharpe-autoinv-validation]] — the
validation.pysketch whosetrial_sharpe_varcomment introduced the conflation this chain has been unwinding - [[2026-08-31-sealed-holdout-refresh-policy]] — the reseal policy whose R3 window-pinning creates the singleton clusters flagged above
- [[2026-05-31-ensembles-systematic-trading-overfitting]] — the False Strategy framing whose σ(trial Sharpes) term is the quantity this brief pins down
Sources
Vault:
~/rdco-vault/06-reference/research/2026-08-31-dsr-effective-n-estimator-precommit.md— step 5 and open-follow-up #1, the subject of this verification~/rdco-vault/06-reference/research/2026-10-29-dsr-cross-block-accounting-after-reseal.md— prior successful primary read (§8.2 quotes, R7, R9'sM_max), and the0.14%/1.1%ratchet sensitivities~/rdco-vault/06-reference/research/2026-09-04-dsr-denominator-sr0-vs-srobs-resolution.md— DSR primary's variable list, "variance across the trials' estimated SR and N is the number of independent trials"~/rdco-vault/06-reference/research/2026-09-05-afml-embargo-cscv-primary-source-verification.md— evidence-tiering precedent~/rdco-vault/06-reference/research/2026-06-01-cpcv-deflated-sharpe-autoinv-validation.md—validation.pysignatures and the original conflation
Web / primary:
- López de Prado, M. and Lewis, M. J., "Detection of False Investment Strategies Using Unsupervised Learning Methods," SSRN 3167017, first version 2018-04-04, version read: 2018-11-01 — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3167017, fetched via the PDF mirror https://smallake.kr/wp-content/uploads/2020/04/SSRN-id3167017.pdf (
curl+ UA spoof,pdftotext -layout, 25 pp., clean extraction). Primary, independently deep-read this run. Load-bearing: §8.2 "reduce the N strategies to K ≪ N cluster-level strategies... for each k = 1,…,K" and the min-variance aggregation + frequency normalization (pp.11-12); §7'sk = 1,…,Kindex coupling (p.8); "The number of independent trials K is less or equal to the number of trials N" (p.9); ONC'sk = 2,…,N−1search range (p.10); blocks of size 1 "difficult to identify as a cluster" (p.12); Appendix A.4 / Snippet 4 inverse-variance + Sherman-Morrison weights (pp.17-18). Caveat: this is the working paper. The published Quantitative Finance 19(9), 1555-1565 (2019) version was NOT read — tandfonline is paywalled and was not retried. - Quantitative Finance 19(9), 1555-1565 bibliographic record, access restricted — https://ideas.repec.org/a/taf/quantf/v19y2019i9p1555-1565.html
- López de Prado, M., "Detection of False Investment Strategies through FWER and FDR" (lecture) — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3799803, linked from https://www.quantresearch.org/Lectures.htm. Not read — SSRN's direct-download endpoint returned a Cloudflare interstitial.
Own computation (reproducible, synthetic data only):
- Monte Carlo comparing
Vover K min-variance cluster representatives againstVover allN_literaltrial Sharpes, with the cluster count in theN_effslot in both arms. T = 1,000 bars, trials generated as√ρ_in·f_k + √(1−ρ_in)·ε_iat 1% bar volatility, 60-300 replications per configuration.SR₀ratios (cluster-representative ÷ all-trials): 1.048 (10×20, ρ_in=0.7, null), 1.141 (20×100, ρ_in=0.95, null), 1.109 (10×20, ρ_in=0.7, one cluster at +0.06%/bar), 1.266 (10×168, ρ_in=0.9, null), 1.353 (unbalanced 400+9×20, ρ_in=0.9, null), 1.528 (same, one sparse cluster at +0.08%/bar). Appendix A.4 inverse-variance weights vs exact min-variance:SR₀ratio 0.935 to 1.017. Run-to-run CV of√V[{ŜR_k}]at K=10: 0.205-0.260. Scripts at/tmp/dsr-sim-1030/(ephemeral; regenerate from this description).