There is no such thing as a correlation-discounted increment: count the retest, recompute N_eff on the pooled ledger, and forbid Eq. (9)
The question
"Decide the cross-block DSR accounting rule after a sealed-holdout reseal: does a candidate re-tested on a later block count as N+1 trials, or does a correlation-discounted increment apply because consecutive blocks share a regime?"
Acronyms used throughout, expanded once here: DSR = Deflated Sharpe Ratio (Bailey and López de Prado 2014). CPCV = Combinatorial Purged Cross-Validation. ONC = Optimal Number of Clusters, the unsupervised clustering algorithm of López de Prado and Lewis. MinTRL = Minimum Track Record Length. ENB = Effective Number of Bets. FWER = familywise error rate. EVT = Extreme Value Theory. N_eff is the correlation-adjusted effective number of trials; N_literal is the raw ledger row count.
This is open-follow-up #2 of [[2026-08-31-sealed-holdout-refresh-policy]], which pre-registered roll-forward reseal and left the cross-block invariant stated but unquantified ("cumulative N spans blocks... or a reseal silently launders the multiple-testing discount"). It composes directly with the estimator pre-registered in [[2026-08-31-dsr-effective-n-estimator-precommit]]. It is a design decision in our pipeline, not a literature summary, so it ends in a rule precise enough to code and pre-register.
What we already know (from the vault)
- The parent policy already fixed the shape; this brief fixes the arithmetic. [[2026-08-31-sealed-holdout-refresh-policy]] pre-registered calendar-triggered roll-forward reseal, a block budget
B = min(5, floor(L_block/252)), a coarsened PASS/FAIL holdout answer, permanent retirement of failed candidates and their param neighborhood, and the invariant that cumulative N spans blocks. It priced the cost of reseal honestly: "a candidate failing block 1 and passing block 2 has been tested twice," and "consecutive blocks are one historical path rather than independent draws, so 'passed three blocks' is materially weaker than three independent trials." - The companion already pre-registered where the correlation discount lives. [[2026-08-31-dsr-effective-n-estimator-precommit]] fixes
N_eff = clamp(ceil(max(K_onc, K_rank, K_mp)), 1, N_literal)computed by clustering the cumulative trial ledger's return matrix, with four adversarial guardrails (pre-registered before code, take the max, clamp to N_literal, refuse below ~30 trials). Its stated reason is the one that governs this brief too: "Every effective-N estimator reduces N, which reduces SR₀, which makes GATE-1 easier to pass... a knob that points in one direction, on a gate whose entire purpose is to be hard to pass." - The denominator question is already closed and does not reopen here. [[2026-09-04-dsr-denominator-sr0-vs-srobs-resolution]] deep-read the DSR primary and confirmed Eq. (2) evaluates the standard error at the observed Sharpe. Nothing in this brief touches that term; the whole cross-block question lives in
SR₀via N. - The data scale is the binding fact and it is verified, not assumed. [[2026-09-05-cpcv-partition-sizing-per-autoinv-universe]] establishes the only cached series as
SPY_2018-01-01_2023-12-31, 1,509 daily bars (2018-01-02 to 2023-12-29), with a train/validation partition of ~1,005 bars. The parent brief's holdout arithmetic runs on ~300 bars.autoinv/validation.pystill contains no holdout, no DSR, no CPCV. - RDCO has run the collapse-correlated-things mechanism twice already ([[2026-06-01-ensemble-admission-correlation-threshold]] at the ensemble layer via Minimum-Linear-Torsion, and position-level bucket caps in [[2026-05-29-strategy-pipeline-architecture-v0]]), with a documented false-negative from the PCA route. Every prior instance was within one sample. Nothing in the vault has ever applied it across disjoint time blocks, which is the new thing here.
What the web says
- The DSR primary does not treat N as a literal backtest count, and it says so in a dedicated appendix. Bailey and López de Prado, Appendix A.3, p.14, verbatim: "It is critical to understand that the N used to compute E[max{SR_n}] corresponds to the number of independent trials. Suppose that we run M trials, where only N trials are independent, N<M. Clearly, using M instead of N will overstate E[max{SR_n}]." The main text is consistent: Eq. (1) is derived for "N independent backtests or track records," and the investor in the worked example asks for "the number of independent trials carried out (N)" (p.7, p.9). So "N is effectively-independent trials" is derived; "N is the literal trial count" is convention, adopted because the honest count is conservative. (Bailey and López de Prado 2014, PDF; SSRN 2460551) Primary, deep-read: full PDF retrieved and pp.14-15 read as rendered page images.
- The primary supplies the correlation-discount formula the question is reaching for, and it is a TOTAL, not an increment. Eq. (8) defines the equal-weighted average correlation
ρ = (ΣᵢΣⱼρᵢⱼ − M)/(M(M−1))over the MxM trial correlation matrix; Eq. (9) isN̂ = ρ̂ + (1 − ρ̂)M(p.14, reproduced from the rendered page). Differentiating,dN̂/dM = (1 − ρ̂), so the "N += 1 − ρ_hat" form in the question is exactly the marginal form of Eq. (9) and is only coherent if the same ρ̂ is simultaneously applied to the entire historical ledger. The "N += 1/(1+ρ)" form appears nowhere in either primary; it is a variance-inflation heuristic imported from elsewhere. - The same primary attaches a sample-size condition to its own formula, and then offers two escapes from it. p.15, verbatim: "in practice M almost always exceeds the sample length, T. Then the estimate of average correlation may itself be overfit. In general for short samples T < ½M(M−1), the correlation matrix will be numerically ill-conditioned... Estimating an average correlation is then pointless, because there are more correlations than independent pairs of observations!" It also flags that "correlation is a limited notion of linear dependence," and that Eq. (9) should be "enriched by incorporating Fisher's transform... thus controlling for the variance of the error in the estimation of ρ̂." Read the rest of the page, because it does not stop at the prohibition. It supplies two remedies: "One way to deal with this numerical problem is to reduce the dimension of the correlation matrix (see Bailey and López de Prado [2012b] for one such algorithm), and compute the average correlation on that reduced definite-positive matrix," and "An alternative and more direct path is to use information theory to determine N̂… data compression, total correlation and multiinformation, e.g. Watabane [1960] and Studený and Vejnarová [1999]." The paper therefore sets a condition, not a permanent bar. Why neither escape helps us is argued in the synthesis below, and that argument is mine, not the paper's.
- The lead author rejected the constant-average-correlation assumption four years later, in the paper written specifically about effective-N — though reading that as a repudiation of Eq. (9) is MY INFERENCE, not the paper's statement. Be precise about what the 2018 paper does and does not say: it aims this passage at Harvey et al. [2015], it never mentions Eq. (9) or the 2014 appendix, and it cites Bailey and López de Prado [2012, 2014] approvingly as the source of the False Strategy theorem whose inputs it is supplying estimators for (p.5). What is quotable is that it declines the assumption Eq. (9) is built on. That the later estimator therefore supersedes the earlier one in practice is my reading and should be weighed as such. López de Prado and Lewis, listing how their method differs from Harvey et al.'s Šidák-under-constant-correlation approach: "We do not assume a constant average correlation across trials. A family of backtests often contains heterogeneous strategies... This clustering of trials around heterogeneous strategies leads to a hierarchical structure... Assuming a constant correlation across all trials fails to recognize that hierarchical structure, biasing the estimates of the number of independent trials (E[K]) and the false positive probability" (p.5-6). The paper's two meta-research variables are E[K], "the number of effectively uncorrelated tests," and E[V[{SR_k}]], "the variance of the SR across the K effectively uncorrelated tests" (p.5). (SSRN 3167017 working paper, Nov 2018 version, read via the smallake.kr PDF mirror) Primary deep-read for the first time in this vault — the known-dead codemacher.com mirror was not re-attempted. Caveat: this is the November 2018 working paper, not the published Quantitative Finance 19(9) version, which is paywalled and unread.
- §8.2 of that paper closes an open follow-up the companion brief left hanging. The within-cluster aggregation rule is minimum-variance weighting:
w_k = Σ_k⁻¹1 / (1'Σ_k⁻¹1),S_k,t = Σ_{i∈C_k} w_k,i·r_i,t; each cluster Sharpe is then annualized at that cluster's own trading frequencyFrequency_k = T_k/Years_kand converted to the selected strategy's frequency,E[V[{SR_k}]] = V[{aSR_k}]/Frequency_k*(pp.11-12). The aggregation rule is primary-confirmed; what follows is not. The paper's stated reason for the normalization is heterogeneous trading frequency — "these SR_k are not yet comparable, as their frequency of trading may vary" — not disjoint time blocks. It is true that the machinery is defined over each cluster's ownFirst Date_k,Last Date_kand lengthT_k, so it is expressible across trials with different spans. Reading that as the primary's answer to the cross-block case is my extension of it, and the paper neither makes nor endorses that claim. - The strongest evidence on the cross-block question comes from outside finance, and it is directly against the discounted-increment route. Halle, Djurovic, Andreassen and Langaas review the Cheverud/Nyholt/Li-Ji/Gao/Galwey family of effective-number-of-tests estimators and conclude, in the abstract: "we show that the effective number of independent tests is not additive over blocks of independent genetic markers unless we assume a common value for the local significance level. We also show that the reviewed methods based on estimating the effective number of independent tests in general do not control the familywise error rate." Their measured example is the direction that matters: for 1,000 markers in 100 blocks, "the method of Cheverud (2001) is not additive as the sum of the block-wise estimates is Meff = 559.0, while the estimated number of independent tests using the whole genotype correlation matrix is Meff = 995.6" (Discussion, p.15). Their conclusion: "The other methods considered in this paper (except the Bonferroni and Šidák methods) fail to control the FWER in at least one of the examples studied." (arXiv:1612.04535v2, stat.ME) Primary, deep-read from full extracted text.
Convergences and contradictions
- Convergence on the direction of the error, from both primaries and from the block-additivity experiment. Bailey and López de Prado state that using the literal M "will overstate E[max{SR_n}]" (p.14), which is the same sentence read as: naive counting produces too high a hurdle, i.e. it is conservative. Halle et al. measure what happens when you build N_eff additively from block-wise pieces instead: 559.0 against a true-matrix 995.6, a ~44% under-count of the trial family, in the permissive direction, on independent blocks where no discount was even warranted. Both agree that the error from naive counting is false negatives and the error from block-wise discounting is false positives.
- Contradiction inside the López de Prado corpus, and the later paper wins. DSR (2014) Appendix A.3 offers Eq. (9) as "one path." López de Prado and Lewis (2018) explicitly decline to "assume a constant average correlation across trials" and say doing so biases E[K] and the false-positive probability. Eq. (9) assumes exactly that. The 2014 appendix is superseded by its own lead author, and the superseding paper is the one the companion pre-registration already builds on.
- The question's framing contains a defect worth naming: there is no coherent "correlation-discounted increment." Eq. (9) is a function of the whole matrix. Applying
N += (1 − ρ̂)to the newest trial while leaving every historical trial at its literal count is internally inconsistent, since the same ρ̂ that justifies discounting trial M+1 also discounts trials 1..M. Applying it consistently means re-opening the entire ledger to a single scalar ρ̂, which is the thing Lewis rejects and the thing the T < ½M(M−1) condition forbids at our scale. Both readings fail, for different reasons. - A second structural problem with cross-block ρ that neither primary addresses. Eq. (8)'s ρᵢⱼ is the correlation of two trials' contemporaneous return series. Under roll-forward reseal, TRAIN grows monotonically, so a block-2 trial's series is a time-superset of its block-1 twin. Their measured correlation is therefore dominated by the shared pre-block-2 history and tends toward 1 by construction, driving the Eq. (9) increment toward 0. The number that comes out is an artifact of the overlap, not a measurement of regime similarity. This is my own analysis, not a cited finding, and it is the single most dangerous thing about the discount route: it would not fail loudly, it would quietly return "this retest cost nothing."
Synthesis for RDCO
The honest answer is the less clever one, and the arithmetic says the clever one buys nothing anyway. Count the retest. N_literal += 1, naive, unconditional. The correlation adjustment is applied exactly once, inside the N_eff clustering that [[2026-08-31-dsr-effective-n-estimator-precommit]] already pre-registered, and it is applied by recomputing N_eff on the pooled cross-block ledger rather than by adding a discounted increment to a running total. Three independent reasons converge on this, and they do not depend on each other. First, the primary's own condition: estimating an average correlation requires T ≥ ½M(M−1). At the ~1,005-bar validation partition that caps the trial matrix at M ≤ 45 trials, and at the ~300-bar block scale at M ≤ 25 (my arithmetic from the primary's condition); even a full-history SPY cache around 8,000 bars caps it at M ≤ 126. v0's sweep counts grid cells in the thousands. The paper does not stop there, and neither should this brief: p.15 offers two escapes from its own condition, and the reason neither rescues Eq. (9) for us is an argument I owe, not a bar the paper sets. The first escape, reduce the dimension of the correlation matrix and compute the average correlation on the reduced definite-positive matrix, does not give us a second discount: collapsing M correlated trials into a smaller set of representatives is the same operation ONC already performs under the companion pre-registration, so taking the reduced-matrix ρ̂ and also clustering would double-count the identical structure. It converges on our rule rather than competing with it. The second escape, information-theoretic estimation of N̂ via total correlation or multiinformation, replaces a second-moment estimator with an entropy estimator over the same M variables and the same T observations; entropy estimation over high-dimensional continuous variables is more sample-hungry than correlation estimation, not less, so at a T where the correlation is already "pointless" by the paper's own test, the information-theoretic route is worse off, not better. That is standard estimation reasoning and it is mine; I did not read Watanabe [1960] or Studený and Vejnarová [1999] and I am not citing them for it. The conclusion survives the correction: the correlation route is unavailable to us at v0's trial counts, by the paper's condition and because its own two remedies do not reach our regime. Second, Halle et al.'s measured non-additivity: summing block-wise effective-N under-counted the family by 44% in a case with no true cross-block dependence, and only the undiscounted Bonferroni and Šidák corrections controlled FWER across all their scenarios. Third, numerical immateriality, below.
State the direction of the error plainly, because this is the self-serving fork. A discounted increment lowers N, which lowers SR₀, which lowers the hurdle, which makes GATE-1 easier to pass. It is the permissive direction, and it is the direction a pipeline operator wants when a candidate is sitting just under the bar. The correlation-discount route would absolutely flip marginal candidates, and the primary's own worked example is the proof. Bailey and López de Prado's example (T=1250, V[{SR_n}]=1/500, γ̂₃=−3, γ̂₄=10, annualized SR 2.5) prints DSR = 0.9004 at N=100, a reject at 95%, and notes that at N=46 it would have been 0.9505, a pass (pp.9-10; both constants reproduce to four decimals in my own computation). Applying Eq. (9) to that same M=100 ledger, ρ̂ = 0.5408 yields N̂ = 46.5 and DSR = 0.9500 — the paper's own rejected strategy becomes a funded one, with no new data, purely from an estimated scalar. A plausible-sounding average correlation of 0.54 among grid-neighbor trials is not a stretch; it is what you would expect. That is the whole risk in one number. By contrast the naive cross-block increment is nearly free: in the same example, moving N from 100 to 101 changes DSR by 0.0008. Across ledger sizes, the gap between N+1 and a half-discounted N+0.5 is 3.9% of SR₀ at M=5, 0.55% at M=20, 0.069% at M=100, and 0.002% at M=2,000 (my computation of the SR₀ formula). The conservative choice costs essentially nothing at any realistic ledger size, and the permissive choice is worth a whole confidence level. There is no trade here worth making.
Can we estimate ρ from our own data? No, and the parent brief's finding generalizes. The cross-block estimand is not even the primary's ρ. It is the correlation, across candidates, between a candidate's block-b Sharpe and its block-(b+1) Sharpe. Estimating it requires at least two realized sealed blocks and a population of candidates tested on both. We have one cached series of 1,509 bars, a queryable-block floor of 1,000 bars under the parent policy, and a minimum reseal accrual of 252 new bars; block 2 does not exist and cannot exist for years. Even granting it: the sample of candidate pairs is itself correlated (grid neighbors), so the effective sample size for ρ̂ is the cluster count, not the candidate count. Under Fisher's z-transform, which the DSR primary itself names as the fix for ρ̂'s estimation error, a true ρ̂ = 0.5 at n = 20 effectively-independent pairs carries a 95% interval of [0.07, 0.77], so the implied increment (1 − ρ̂) ranges over [0.23, 0.93] — a four-fold range on the quantity doing the discounting. At n = 10 the interval includes zero. This is the same finding the parent brief reached about the holdout Sharpe itself: estimator noise, not leakage, is what binds. A rule requiring a well-estimated ρ is unimplementable at RDCO's data scale, and that is the finding, not a deferral.
Composition with the companion estimator: one discount, in one place, with a ratchet on the hurdle itself. The cross-block correlation adjustment and the within-block clustering adjustment are the same adjustment if, and only if, N_eff is recomputed from scratch on the pooled ledger rather than accumulated. When a block-2 sweep rediscovers a block-1 candidate, ONC will place the two rows in one cluster and the pair will contribute roughly 1 to N_eff rather than 2. That is the discount, applied once. Adding N += (1 − ρ̂) on top is the double-discount the question warns about. There is a second, subtler double-count to avoid in V[{SR_n}]: for M equicorrelated trials the expected sample cross-sectional variance is σ²(1 − ρ), not σ² (standard result, derived here), so computing V[{SR_n}] across all correlated trial Sharpes already shrinks the dispersion by (1 − ρ) and shrinks SR₀ by √(1 − ρ) — a 29% lower hurdle at ρ = 0.5, entirely silently. López de Prado and Lewis §8.2 pre-empts this by taking the variance across the K cluster-representative series, which is what the companion brief's step 5 already specifies. That step is now primary-confirmed, including its aggregation rule (minimum-variance weights, annualize per cluster frequency, express at the selected strategy's frequency), closing the companion's open follow-up #1 (the aggregation rule is what the primary confirms; the cross-block reading of it is mine). Keep it; do not substitute the all-trials dispersion, which is the permissive one. And close the loop, because naming the variance channel is not the same as guarding it: since SR₀ is linear in √V[{SR_k}], every one of these dispersion effects is a hurdle effect, and an anti-laundering rule that polices only the trial count cannot see them. A 0.14% drop in the cluster-Sharpe dispersion at N_eff = 100 cancels a whole extra trial; 1.1% does it at N_eff = 20 (my arithmetic). That is why R4 below asserts monotonicity on SR₀ and not on N_eff, which is the single most important change between this rule and the obvious version of it.
PRE-REGISTERED RULE — CROSS-BLOCK TRIAL ACCOUNTING (CBTA-1). This is the thing to code.
A candidate re-tested on a later block counts as a full additional trial in the ledger. No correlation-discounted increment is ever applied. The correlation adjustment happens once, by recomputing N_eff on the pooled cross-block ledger.
R1 — Ledger counter. Every evaluation of a candidate against a new block's validation partition writes a new ledger row:
N_literal += 1, unconditionally, regardless of similarity to any prior row. The row recordsblock_id,prior_trial_id(the earlier row it re-tests, or null), andretest_class ∈ {NEW, REDISCOVERY, RECONFIRM}, assigned before the trial is scored.R2 — One discount, one place.
N_effis recomputed from scratch on the pooled cross-block ledger per [[2026-08-31-dsr-effective-n-estimator-precommit]] steps 2-4. It is never accumulated from per-block estimates, and no ρ term appears anywhere else in the pipeline. Rationale is citable, not stylistic: block-wise effective-N is not additive (Halle et al. 2016).R3 — Pinned clustering window (anti-laundering). The correlation matrix
Cfed to ONC is computed on a single pinned reference windowW_ref, with every surviving ledger trial's return series re-evaluated onW_ref(cheap: re-run the param vector through the vectorbt sweep).W_refis the current block epoch's TRAIN/VALIDATION partition — about 1,005 bars today and growing with each reseal — explicitly NOT the ~300-bar sealed holdout block, which stays import-walled out of this object entirely. Using each trial's own full history instead would let shared pre-block data manufacture correlation and silently shrinkN_eff. A trial that cannot be re-evaluated onW_refkeeps its ledger row, is excluded from the clustering, and counts as its own cluster (+1 toN_eff).R4 — Monotonicity invariant, asserted on the hurdle and not on the count. The quantity that must never fall across a reseal is
SR₀, notN_eff.SR₀ = √V[{SR_k}] · ((1−γ)Φ⁻¹(1−1/N_eff) + γΦ⁻¹(1−1/(N_eff·e)))is linear in the cluster-Sharpe dispersion, so a monotoneN_effdoes not imply a monotone hurdle: atN_eff = 100a 0.14% fall in√V[{SR_k}]exactly cancelsN_eff += 1, and atN_eff = 20it takes 1.1% (my arithmetic from the SR₀ formula). An accounting that polices only the count leaves the whole variance channel open, in the permissive direction. Therefore:SR₀_gate = max(SR₀_new, SR₀_prev), compared in the companion's pinned units (per-return-observation, same bar frequency, same cost model). LogN_eff,N_eff_prev,V[{SR_k}],V_prev,SR₀_new,SR₀_prev, andSR₀_gateon every post-reseal run.R4a — Resolution of the R2/R3/R4 conflict, pre-registered rather than left to the implementer. R2 recomputes
N_efffrom scratch and R3 pins a newW_refeach block, so neither carries a monotonicity guarantee and a legitimate run can produceSR₀_new < SR₀_prev. R4 must therefore clamp and flag, not halt: takeSR₀_gate = max(SR₀_new, SR₀_prev), mark the runHURDLE_RATCHET_APPLIED, and surface it for review. A hard halt would create pressure to reuse the previous window, which silently violates R3 and is exactly the convenient wrong answer. Failing conservative beats failing closed here.R5 — RECONFIRM carve-out (N_eff contribution = 0). Re-testing an already-DEPLOYED, previously-PASSED candidate on a new block writes
N_literal += 1for audit but is excluded from the GATE-1 family only if all four hold: (a) the bar was pre-registered before the block sealed; (b) no parameter was re-tuned, at all; (c) the test is two-sided with a pre-committed DE-DEPLOY action on fail, so it is not a free option; (d) the row cites the original pass'strial_id. Any violation reclassifies the row to REDISCOVERY and it re-enters the family. Rationale: a single pre-specified hypothesis with a symmetric consequence is not a selection among alternatives, so it contributes no selection bias to the discovery family. This is derived, not cited — medium confidence, and the four conditions are what make it defensible. Direction of error, stated because it is the exception: R5 is the only clause in CBTA-1 that lowers a hurdle, so it is the only one that can produce a false positive. The four conditions are conjunctive precisely to bound it, and (b) and (c) are the load-bearing pair — a re-tuned parameter makes it a selection, and a fail with no de-deploy makes it a free option. If a reviewer has to argue about whether a condition is met, it is not met.R6 — Resurrection stays forbidden. [[2026-08-31-sealed-holdout-refresh-policy]] permanently retires a FAILED candidate and its param neighborhood. CBTA-1 does not relax that. If it is ever relaxed, the resurrected candidate is excluded from cluster-collapse and floors
N_effat +1 with no discount, because a second attempt at the same gate is pure alpha-spending.R7 —
V[{SR_k}]stays on cluster representatives. Per López de Prado and Lewis §8.2: min-variance-weighted cluster return series, each annualized at its ownFrequency_k, expressed at the selected strategy's frequency. Log the all-trials dispersion alongside it for audit, flagged not-for-use. If the cluster-representative value is ever the smaller of the two, flag the run for review rather than proceeding.R8 — Eq. (9) is forbidden, with a visible reason. No average-correlation discount may enter
SR₀. Reinstating it requiresT ≥ ½M(M−1)on the clustering window, per the primary (p.15). The ledger reportsM_max = floor((1+√(1+8·T_ref))/2)next toN_effon every run, computed onW_ref's length, so the violation is visible in the artifact rather than argued about later.R9 — ONC gets NO exemption from the condition used to forbid Eq. (9), and this is an amendment to the companion's refusal clause. It would be incoherent to forbid the average-correlation route on
T < ½M(M−1)while feeding the same MxM matrix to ONC unguarded. The failure mode is real and permissive: withM > Tthe sample correlation matrix has rank at mostT, so spurious shared structure appears, clusters over-collapse,K_oncfalls,N_efffalls, and the hurdle falls. [[2026-08-31-dsr-effective-n-estimator-precommit]] already refuses and falls back toN_eff = N_literalwhen the trial matrix is "degenerate"; that word is hereby given a number. Degenerate meansM > M_max(T_ref). When it holds, the estimator refuses:N_eff = N_literal, no discount, run flaggedCLUSTERING_REFUSED_ILLCONDITIONED. The stricter½M(M−1)bar is used rather than the looser rank barM ≤ Tbecause more refusals means a higher hurdle, which is the direction this gate is supposed to err in.What R9 actually costs, said plainly rather than buried. At
T_ref ≈ 1,005bars,M_max = 45. v0's sweep counts grid cells in the thousands, so under CBTA-1 as written the clustering discount will refuse andN_effwill equalN_literalon every realistic v0 run. TheN_effmachinery is not dead, but it is dormant until one of two things changes: the sweep keepsMunderM_max(which is v0's original blunt mitigation, "cap grid granularity so cells are meaningfully distinct," arrived at here from the opposite direction and now with a number attached), or the ledger is dimension-reduced before clustering per the primary's own first remedy. Either is a legitimate way to earn the discount. Neither is a reason to skip the guard.Report per run:
N_literal,N_eff,N_eff_prev, the three cluster estimates, the per-block row counts byretest_class,W_refstart/end andT_ref, the count of trials excluded from clustering under R3,M,M_max,V[{SR_k}]both ways plusV_prev,SR₀_new,SR₀_prev,SR₀_gate, and theHURDLE_RATCHET_APPLIED/CLUSTERING_REFUSED_ILLCONDITIONEDflags, alongsideSR_obsandT.
Test and implementation plan, so this is checkable rather than asserted. (1) Unit test R4 on the hurdle, not the count: construct a two-block synthetic ledger in which N_eff rises by 1 while the cluster-Sharpe dispersion falls by 0.2% at N_eff ≈ 100, confirm SR₀_new < SR₀_prev (the laundering channel a count-only invariant would miss), and assert the ratchet fires and sets HURDLE_RATCHET_APPLIED. Separately assert the pooled-history clustering window fails loudly against R3. (1a) Unit test R9: feed a matrix with M > M_max(T_ref) and assert N_eff == N_literal with the refusal flag set, rather than a silently-collapsed K_onc. (2) Unit test R3's artifact: generate one strategy evaluated on a short window and the same params on a time-superset window, confirm their measured correlation on the pooled-history construction exceeds their correlation on W_ref, which is the mechanism R3 defends against. (3) Regression-test the DSR implementation against the primary's printed constants before anything else is trusted: N=100 must give 0.9004 and N=46 must give 0.9505 on the paper's inputs. (4) Extend v0's pure-noise 100-cell V-model test across a synthetic reseal and assert the false-positive rate does not rise after the second block. (5) Log retest_class from the first block, before any reseal exists, so the classification is never retrofitted.
Why this is in the vault
This closes open-follow-up #2 of [[2026-08-31-sealed-holdout-refresh-policy]] with a rule that can be coded as written, and it is a Phase-2 build input to two specific unbuilt objects in /Users/ray/Projects/automated-investing: the HoldoutManager reseal path and the multiple-testing-ledger schema behind validation.py: deflated_sharpe(...). It also closes open-follow-up #1 of [[2026-08-31-dsr-effective-n-estimator-precommit]] as a byproduct, by reaching the López de Prado and Lewis primary (§8.2, minimum-variance cluster aggregation) that two prior attempts failed to extract. Concretely, it prevents the specific failure where the first reseal silently launders the trial discount and every DSR the pipeline reports afterward is too permissive.
Open follow-ups
- Does the Extreme Value Theory derivation survive the substitution of
N_efffor N? Eq. (5)-(6) of the DSR primary are proved for independent and identically distributed standard normals (Appendix A.1, p.12). Plugging a correlation-adjustedN_effinto a formula derived under independence is an approximation whose error is quantified nowhere in either primary. The dependent-extremes literature (the extremal index, and Halle et al.'s "Order 2" method, the only one that controlled FWER in all their scenarios) bears directly on it. Neither parent brief raises this; the companion assumes the substitution is valid, and this brief inherits that assumption. - Would a resampling-based
SR₀remove the need to estimate dependence at all? A Westfall-Young-style max-statistic null, built by stationary block bootstrap over the trial-by-time return matrix we already have to store, respects the dependence structure exactly instead of summarizing it in a scalar. The research question is whether that is calibrated atT ≈ 1,000bars and what block length the literature supports for Sharpe-ratio statistics. Neither parent brief mentions resampling anywhere. - What does Harvey, Liu and Zhu's framework say about the same cross-block question? The DSR primary names Harvey and Liu as the leading alternative (p.9, "The role of HL's threshold is analogous to the role played by our E[max{SR_n}]"), and López de Prado and Lewis contrast with Harvey et al. [2015] explicitly. The HL "p-value haircut" line is a genuine rival framework for multiple testing in finance and has never been read into this vault. It may give a different, Benjamini-Hochberg-flavored answer to repeated testing across blocks.
- Is there a published stability condition for ONC's K? The companion pre-registered a ≥30-trial fallback floor with no empirical basis. The November 2018 working paper I read states no minimum trial count or minimum T for the clustering to be stable; whether the published Quantitative Finance 19(9) 2019 version adds one is unknown and is the kind of condition that would change R3 and the fallback rule.
- Is regime similarity between consecutive blocks measurable directly, rather than inferred from trial-Sharpe correlation? The question's premise is "consecutive blocks share a regime." The distribution-shift literature (dataset distance, test-set similarity measures) offers direct statistics on whether a new test block is a genuinely new sample. If one of those is estimable on ~300 bars, it could gate whether a reseal counts as a fresh test surface at all — a cleaner lever than discounting N, and one that errs conservative when the blocks are similar.
Related
- [[2026-08-31-sealed-holdout-refresh-policy]] — the parent brief whose open-follow-up #2 this closes; reseal cadence, block budget B, permanent retirement, and the cross-block cumulative-N invariant this brief makes operational
- [[2026-08-31-dsr-effective-n-estimator-precommit]] — the companion pre-registration this composes with; its
N_effclustering is the single place the correlation discount is applied, and its step-5 cluster-representativeV[{SR_n}]aggregation rule is primary-confirmed here, while R9 below puts a number on its "degenerate matrix" refusal clause - [[2026-09-04-dsr-denominator-sr0-vs-srobs-resolution]] — fixes the DSR denominator at
SR_obs; untouched by this brief, which only affectsSR₀via N - [[2026-09-05-cpcv-partition-sizing-per-autoinv-universe]] — source of the verified ~1,005-bar train/validation partition and 1,509-bar cache that set the
M ≤ 45ceiling - [[2026-06-01-cpcv-deflated-sharpe-autoinv-validation]] — DSR / SR₀ / MinTRL formulas and the
validation.pysketch the ledger schema extends - [[2026-05-29-strategy-pipeline-architecture-v0]] — GATE-0/1/2 design, the cumulative auto-N trial ledger, and the pure-noise V-model test extended in the plan above
- [[2026-06-01-ensemble-admission-correlation-threshold]] — RDCO's existing correlation-to-independence machinery (ENB via Minimum-Linear-Torsion) and its documented PCA false-negative
- [[2026-05-31-ensembles-systematic-trading-overfitting]] — the False Strategy framing
E[max SR] ≈ √(2·ln N)·σ(trial Sharpes)that makes the ln-N flatness argument above concrete
Sources
Vault:
~/rdco-vault/06-reference/research/2026-08-31-sealed-holdout-refresh-policy.md— reseal policy, open-follow-up #2 verbatim, cross-block cumulative-N invariant, ~300-bar holdout arithmetic~/rdco-vault/06-reference/research/2026-08-31-dsr-effective-n-estimator-precommit.md— pre-registeredN_effestimator, four adversarial guardrails, open-follow-up #1 closed here~/rdco-vault/06-reference/research/2026-09-04-dsr-denominator-sr0-vs-srobs-resolution.md— DSR Eq. (2) denominator atSR_obs; page-verified read of the same primary~/rdco-vault/06-reference/research/2026-09-05-cpcv-partition-sizing-per-autoinv-universe.md— 1,509 cached bars, ~1,005-bar train/validation partition, repo state~/rdco-vault/06-reference/research/2026-06-01-cpcv-deflated-sharpe-autoinv-validation.md— DSR/SR₀/MinTRL formulas,validation.pysketch~/rdco-vault/01-projects/investing/2026-05-29-strategy-pipeline-architecture-v0.md— GATE-0/1/2, cumulative trial ledger, pure-noise V-model test~/rdco-vault/06-reference/research/2026-06-01-ensemble-admission-correlation-threshold.md— ENB via Minimum-Linear-Torsion, PCA false-negative~/rdco-vault/06-reference/research/2026-05-31-ensembles-systematic-trading-overfitting.md— False Strategy Theorem framing
Web (all three primaries below were deep-read from retrieved full text, not search extracts):
- Bailey, D. H. and López de Prado, M., "The Deflated Sharpe Ratio," Journal of Portfolio Management, 2014 — https://www.davidhbailey.com/dhbpapers/deflated-sharpe.pdf and https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2460551. Primary, deep-read: full PDF retrieved, text extracted, and pp.14-15 read as rendered page images to recover the equations. Load-bearing citations: N as "the number of independent trials" and "using M instead of N will overstate E[max{SR_n}]" (A.3, p.14); Eq. (8) average correlation and Eq. (9)
N̂ = ρ̂ + (1−ρ̂)M(p.14); theT < ½M(M−1)ill-conditioning condition and "estimating an average correlation is then pointless" (p.15); Fisher's-transform caveat (p.14); iid-normal Extreme Value Theory derivation (A.1, p.12); worked example N=100 → DSR 0.9004, N=46 → 0.9505 (pp.9-10). - López de Prado, M. and Lewis, M. J., "Detection of False Investment Strategies Using Unsupervised Learning Methods," SSRN working paper, first version 2018-04-04, version read: 2018-11-01 — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3167017, via PDF mirror https://smallake.kr/wp-content/uploads/2020/04/SSRN-id3167017.pdf. Primary, deep-read — first successful extraction in this vault. Load-bearing: the two meta-research variables E[K] and E[V[{SR_k}]] (p.5); "We do not assume a constant average correlation across trials... biasing the estimates of the number of independent trials" (pp.5-6); §8.1 ONC; §8.2 minimum-variance cluster aggregation and frequency normalization (pp.11-12). Caveat: this is the working paper. The published Quantitative Finance 19(9), 1555-1565 (2019) version is paywalled on tandfonline and was NOT read; differences are unverified. The previously-dead codemacher.com mirror was not re-attempted.
- Halle, K. K., Djurovic, S., Andreassen, O. A. and Langaas, M., "Is the familywise error rate in genomics controlled by methods based on the effective number of independent tests?" arXiv:1612.04535v2 [stat.ME], 2016 — https://arxiv.org/pdf/1612.04535. Primary, deep-read from full extracted text. Load-bearing: non-additivity of effective-N over blocks and general failure to control FWER (abstract); Cheverud block-sum Meff = 559.0 vs whole-matrix Meff = 995.6 (Discussion, p.15); "except the Bonferroni and Šidák methods" (Conclusion, p.15). Note: this is the genomics literature, and transfer to backtest trials is by analogy of structure (correlated test statistics, block-organized), not by domain.
Computed in this brief (my own arithmetic, from the primaries' published formulas and inputs; reproducible from the formulas above):
- Reproduction of the DSR primary's printed constants: N=46 → DSR 0.9505, N=100 → DSR 0.9004, both to four decimals. The implementation that reproduces them is the one used for every other number here.
- The Eq. (9) flip point on the paper's own example: ρ̂ = 0.5408 → N̂ = 46.5 → DSR = 0.9500, converting the paper's rejected candidate into a 95%-confidence pass.
- Marginal cost of naive counting, as a fraction of SR₀, for
N+1vsN+0.5: 3.9% at M=5, 0.55% at M=20, 0.069% at M=100, 0.002% at M=2,000. In DSR terms on the paper's example, N=100 → 101 costs 0.0008. - The variance channel that R4 closes:
SR₀is linear in the cluster-Sharpe dispersion√V[{SR_k}], so the dispersion fall that exactly cancelsN_eff += 1is 0.138% at N_eff = 100 (SR₀/√V moves 2.5306 → 2.53410) and 1.11% at N_eff = 20 (1.9007 → 1.92211). A monotone-N_effinvariant does not see either. - The primary's
T ≥ ½M(M−1)condition solved for M: T=300 → M≤25; T=1,005 → M≤45; T=1,509 → M≤55; T=8,000 → M≤126. - Fisher-z interval on ρ̂ = 0.5: n=10 → [−0.19, 0.86]; n=20 → [0.07, 0.77]; n=30 → [0.17, 0.73]. Implied
(1−ρ̂)increment at n=20 spans [0.23, 0.93]. E[S²] = σ²(1−ρ)for M equicorrelated trial Sharpes (standard result, derived here), implyingSR₀is scaled by√(1−ρ)whenV[{SR_n}]is taken over the correlated trial set: 0.949 at ρ=0.1, 0.707 at ρ=0.5, 0.548 at ρ=0.7.