At a 0.5 Sharpe gap, L_min is 8x to 11x the policy's 1,000 bars; nearly all of that is plain iid noise, and the small non-iid remainder is variance dependence rather than autocorrelation or unconditional fat tails
The question
"Verify the Lo (2002) Sharpe standard-error arithmetic under autocorrelated, fat-tailed daily returns rather than the iid approximation — how much longer than 1,000 bars does L_min actually need to be?"
This closes open-follow-up #4 of [[2026-08-31-sealed-holdout-refresh-policy]], which set the sealed-holdout queryability threshold at L_min = 1,000 daily bars and labelled its own supporting arithmetic "an optimistic floor" because it assumed independent and identically distributed (iid) returns.
What we already know (from the vault)
- [[2026-08-31-sealed-holdout-refresh-policy]] carries the figure being challenged: under the iid approximation the standard error of an annualized Sharpe on T daily bars is about sqrt(252/T), so a 300-bar holdout gives roughly +/-0.92 and a +/-0.35 standard error needs roughly 2,060 bars. It concluded the binding constraint is estimator noise rather than query leakage. Its caveat is worth quoting exactly, because this brief confirms rather than overturns it: "that is my own arithmetic under an iid assumption financial returns violate, so treat it as an optimistic floor." It did not name which violation would dominate.
- That brief pre-registered four parameters, of which two matter here: L_min = 1,000 daily bars for queryability (inside its parameter 2), and a coarsened verdict that returns "PASS/FAIL against the pre-registered three-bar wet-rehearsal criterion plus an interval" and never a rankable score (its parameter 4). It also proposed cross-sectional sealing as the fast way to buy holdout capacity, with an "effective-independence discount" it left unspecified.
- The wet-rehearsal bar is already benchmark-relative, which changes what this brief can usefully recommend. [[2026-05-29-markov-system-requirements-v0]] line 123, Bar 3: "Compare the vol-overlaid book vs the un-overlaid buy-and-hold book ... To turn ON: materially higher Sharpe AND materially lower max drawdown, NET of estimated transaction costs." So the holdout query was never an absolute-Sharpe test. What "materially" means was never defined, and that undefined word is exactly the hole the arithmetic below fills.
- [[2026-06-01-cpcv-deflated-sharpe-autoinv-validation]] establishes that Minimum Track Record Length (MinTRL) and the Deflated Sharpe Ratio (DSR) are defined over T = return observations, and that below roughly 30 near-independent observations the skew and kurtosis feeding the deflation term are estimation noise. That caveat turns out to be the central finding here, not a footnote.
- [[2026-07-05-autoinv-cpcv-vs-walk-forward-confirm-stage]] locked the confirm stage as anchored walk-forward plus Combinatorially Purged Cross-Validation (CPCV), so the holdout is not carrying the statistical inference load. That remains the most important mitigating fact in this whole analysis.
- Verified against the data, not the notes (2026-10-29). The pipeline's only cached series is
SPY_2018-01-01_2023-12-31_5f6c0c96.csv: 1,509 price rows, 1,508 log returns, 5.98 years. Measured by us: SPY's own realized annualized Sharpe over that window is 0.548. That is the scale of the quantity the holdout is being asked to adjudicate.
What the web says
- Lo's iid formula is confirmed exactly against the primary. Lo (2002) Eq 8 and Eq 9 give V_IID = 1 + SR^2/2 and SE(SR_hat) = sqrt((1 + SR^2/2)/T), where SR is the per-period Sharpe, not an annualized one, and T the number of per-period observations. Annualization is a separate equation, Eq 18: V_IID(q) = q(1 + SR^2/2), so SE[SR_hat(q)] = sqrt(q(1 + SR^2/2)/T) with SR still per-period (Lo 2002, Financial Analysts Journal 58(4), pp. 36-52). The risk-free rate is a known constant throughout and contributes no estimation error. One nuance: Lo attributes Eq 8 to iid alone, but his Eq A33 reaches it only after setting nu_3 = 0 and nu_4 = 3*sigma^4, so the result is iid normal.
- For serially dependent returns Lo gives a general Generalized Method of Moments (GMM) form plus an aggregation factor, and no closed-form first-order autoregressive, AR(1), standard-error multiplier. Eq 14 is V_GMM = (dg/dtheta) Sigma (dg/dtheta'), with Sigma the long-run covariance of the moment conditions phi = (R_t - mu, (R_t - mu)^2 - sigma^2), estimated by Newey-West (Eq A12-A14, Bartlett kernel, truncation lag m). Eq 20 is the aggregation factor SR(q) = eta(q) SR with eta(q) = q / sqrt(q + 2 sum_{k=1}^{q-1} (q-k) rho_k), and Eq 22 its AR(1) closed form. Lo states the direction explicitly: positive serial correlation pushes eta(q) below the iid sqrt(q), negative pushes it above.
- Lo's general form already covers volatility clustering in principle, which narrows what is actually broken. Because the second moment condition is (R_t - mu)^2 - sigma^2, its long-run variance is the sum of autocovariances of squared deviations, which is precisely the dependence a Generalized Autoregressive Conditional Heteroskedasticity (GARCH) process generates. So a correctly-specified Newey-West estimate of Sigma captures variance clustering asymptotically whenever the relevant moments exist. What fails is not Lo's framework but the closed forms people actually use: his AR(1) example, which models dependence only in the level of returns, and the Mertens plug-in below.
- Lo warns against the robust estimator unless autocorrelation is actually present. Eq 27 prices the efficiency cost of the time-aggregated robust estimator when returns are in fact iid, and his Table 3 shows the asymptotic variance reaching 334 percent of V_IID at a monthly Sharpe of 1.00 and q=12. His conclusion, verbatim: "unless there is significant serial correlation in return series R_t, the robust Sharpe ratio estimator should not be used." He recommends the Ljung-Box statistic (Eq 28) as the pre-test. His own empirical correction is to the point estimate: a mortgage-backed-securities fund's annual Sharpe falls from 4.03 naive to 2.44 robust, an overstatement of 65 percent, while a risk-arbitrage fund and Washington Mutual Investors both move up because their monthly returns are negatively autocorrelated.
- The non-normal correction is Mertens's, and it is already latent in Lo. The standard non-normal iid Sharpe variance is Var(SR_hat) = (1/T)[1 + SR^2/2 - SR*g3 + SR^2(g4-3)/4], with g3 = skewness and g4 = raw kurtosis. Two independent reputable secondaries state it identically including the negative skewness sign: SLCG's Sharpe Ratio Practice Note and Bailey & Lopez de Prado, "The Sharpe Ratio Efficient Frontier" Eq 8, which writes the algebraically identical collapsed form 1 - g3*SR + SR^2(g4-1)/4 one line later and uses it in the Probabilistic and Deflated Sharpe Ratios. Mertens (2002) itself is unreachable: absent from the author's site, 404 at the circulated URL, no Wayback snapshots. Setting q=1 in Lo's Eq A32 recovers the same expression, so the mathematics traces to a primary even though Mertens's note does not; Lo never isolates, names, or applies it.
- Daily US index autocorrelation flipped sign in 1999 and is now negative. Baltussen, van Bekkum and Da, Journal of Financial Economics 132(1) 2019 Table 1 reports S&P 500 daily AR(1) of +0.103 (t=7.45) before 3 March 1999 and -0.076 (t=-3.29) after, difference significant at t=-6.64, with the flip holding for 16 of 20 world indexes. They attribute it to the growth of index products and arbitrage against the underlying, and note serial dependence becomes indistinguishable from zero once a one-day implementation lag removes nonsynchronous-trading effects.
- Ledoit & Wolf target the Sharpe difference and say no simple formula exists for the serially dependent single-Sharpe case. Their method is a studentized circular-block bootstrap for Delta = Sh_1 - Sh_2 on contemporaneous return pairs (Journal of Empirical Finance 15 (2008) 850-859). Remark 3.1 classifies Lo's iid section as assuming iid normal data and his non-IID section as heteroskedasticity- and autocorrelation-consistent (HAC) inference, and rejects Opdyke's (2007) stationary-case formulae precisely because they reduce to the iid ones: "in the case of general stationary data, simple and easily-implemented formulae do not exist." Bailey & Lopez de Prado read the same identity the opposite way, as evidence Mertens is more general than claimed. This is a live disagreement; Opdyke (2007) is paywalled at Springer and was not retrieved, so it is unadjudicated here.
Convergences and contradictions
- The parent brief's iid arithmetic reproduces exactly, and the check was worth running. Computed by us from Lo Eq 18: sqrt(252/300) = 0.9165, and the full formula at an annualized Sharpe of 1.0 (daily SR = 0.06299) gives 0.9174. The parent's sqrt(252/T) shortcut drops Lo's SR^2/2 term, worth 0.099 percent at daily frequency. +/-0.92 confirmed to two decimals, and 2,060 bars for a +/-0.35 standard error confirmed (we compute 2,057 at SR=0, 2,061 at SR=1.0). The shortcut is sound for a reason the parent did not state: the correction term scales with the per-period Sharpe, which is tiny at daily frequency.
- Of the three candidate iid violations, two run the wrong way for the pessimistic case and the third is real but modest. (i) Autocorrelation is a tailwind on this surface. Our derivation inside Lo's GMM frame gives V_AR1 = (1+rho)/(1-rho) + (SR^2/2)(1+rho^2)/(1-rho^2), collapsing to Lo Eq 8 at rho=0; Monte Carlo over 40,000 paths confirms it within 1 percent at every T and rho tested. At the post-1999 rho of -0.076 the standard error is 0.93x the iid value and L_min is 0.86x, 14 percent shorter. It would have been a headwind on pre-1999 data (rho=+0.103 gives L_min x1.23 for the AR(1) leg alone). (ii) Unconditional fat tails barely matter at daily frequency. Every Mertens term carries a factor of the per-period Sharpe or its square, which at daily frequency are 0.063 and 0.004. Computed by us at an annualized Sharpe of 1.0, the standard-error inflation is 1.8 percent for an ordinary window (g3=-0.51, g4=7.5, illustrative) and 9.6 percent even for a window containing the 1987 crash (g3=-2.39, g4=53.6, unverified secondary); the same scenarios at monthly frequency give 11.0 and 63.6 percent. The Mertens correction is a monthly-frequency problem misapplied to a daily-frequency pipeline. (iii) Variance dependence is the one that bites, it is neither of the above, and on parameters fitted to our own series it is worth only 1.04x to 1.37x on L_min.
- The pipeline's own data is strongly variance-dependent, and at its fitted parameters the fourth moment sits at or beyond the boundary of existence. Ljung-Box(5) on the cached SPY returns is 55.5 on the full window but only 2.99 excluding a single 50-day COVID window (critical value 11.07), so there is no robust evidence of level autocorrelation to correct. Ljung-Box(5) on squared returns is 1,311 full and 118.6 ex-COVID: variance clustering is overwhelming and survives COVID removal. Fitting GARCH(1,1) by Gaussian quasi-maximum-likelihood with variance targeting to our own 1,508 bars gives alpha=0.192, beta=0.778 on the full window and alpha=0.148, beta=0.814 ex-COVID, with standardized-residual kurtosis near 4.8 (implied Student-t degrees of freedom about 7.2). Applying the GARCH(1,1) fourth-moment condition kalpha^2 + 2alpha*beta + beta^2 < 1, the two windows give 1.082 and 1.011. The full-window value is clearly beyond the boundary; the ex-COVID value is a 1,508-bar point estimate within estimation error of 1, so the honest reading is "at or beyond the boundary" rather than definitively infinite. Either way a plug-in standard error that estimates kurtosis has no secure asymptotic justification on this series. The empirical instability is visible directly: raw kurtosis is 14.9 full window and 5.6 ex-COVID, a factor of 2.7 from one 50-day exclusion. Our Monte Carlo also shows the standard error exceeding the AR(1)-plus-Mertens prediction by an amount that grows across the tested range (7 percent at T=300, 13 percent at T=1,000, 20 percent at T=2,520), with the decomposition placing all of the inflation in the variance estimator and none in the mean estimator. We cannot cleanly separate "fourth moment infinite" from "finite but converging slowly near the boundary": the Gaussian-innovation case at alpha=0.09/beta=0.90 satisfies the condition at 0.9963 yet its ratio still climbs from 1.04 to 1.17 out to T=10,080, while a comfortably-finite case (alpha=0.05/beta=0.90) settles at 1.01.
Synthesis for RDCO
The answer to the question asked, simulated at the GARCH parameters fitted to our own cached SPY rather than at assumed ones. Across the four data-fitted variants (full-window and ex-COVID fits, each with the measured and the literature rho), the Monte Carlo standard-error curves run from SE = 13.70 * T^-0.4810 to 11.52 * T^-0.4483 against the iid 15.94 * T^-0.4999. Solving for 95 percent two-sided significance and 80 percent power, L_min to resolve an annualized-Sharpe gap of 0.50 is 8,300 to 10,910 bars, 32.9 to 43.3 years, i.e. 8.3x to 10.9x the policy's 1,000 bars. An earlier version of this brief put the range at 13x to 24x by simulating assumed parameters at GARCH persistence 0.99; our own fits come out at 0.962 and 0.970, and persistence is the main driver, so those figures belong in a sensitivity row rather than the headline. Held as sensitivity: persistence 0.99 at alpha=0.09/beta=0.90 gives 12,967 bars (13.0x) and at alpha=0.15/beta=0.84 gives 24,682 bars (24.7x). The decomposition matters more than the multiple. On a log scale 87 to 98 percent of the gap is plain iid sampling noise: the iid requirement is already 7,985 bars, 8.0x the policy's L_min, before any dependence is modelled. The departure from iid contributes only 1.04x to 1.37x on fitted parameters (1.62x to 3.09x on the higher-persistence sensitivity cases), and that is net of an AR(1) leg running the other way at 0.86x. So volatility clustering is the culprit for the departure from iid, and the departure is modest; the gap itself is overwhelmingly the iid arithmetic the parent brief already had. The parent's "optimistic floor" label was correct, and its stated reason was correct too: variance dependence is itself an iid violation, so nothing it wrote needs retracting. What this brief adds is which violation dominates, that the dominant one is invisible to both corrections the follow-up named, and that it is smaller than the headline framing implied. The choice of a 0.50 gap is ours, not the parent's: across d in {0.25, 0.50, 0.75, 1.00, 1.50} the multiple runs 35x to 51x, 8.3x to 10.9x, 3.6x to 4.4x, 2.0x to 2.3x, and 0.8x to 0.9x. Two honesty notes: this is a Monte-Carlo-fitted result, not a closed form, and the simulation grid stops at T=10,080, so the 0.50 answer sits at or just beyond the grid edge while the 0.25 answer extrapolates several-fold past it.
Turn it around and the indictment is sharper than any L_min number, and it does not depend on the GARCH parameterization at all. At the policy's 1,000-bar L_min the smallest annualized-Sharpe gap resolvable at 95/80 is 1.38 to 1.46 on the fitted specs, and 1.41 even under pure iid. SPY's own realized Sharpe over the cached window is 0.548. The holdout at its own queryability threshold can only detect an edge roughly 2.5x to 2.7x the size of the entire thing being traded. Across the whole 1,500-bar cache it is 1.14 to 1.22, still 2.1x to 2.2x SPY's Sharpe. Because the iid leg dominates, tightening the dependence model cannot rescue this. The undefined "materially higher Sharpe" in wet-rehearsal Bar 3 therefore has an implied floor nobody chose: whatever the holdout can resolve, which is a gap larger than the benchmark's own Sharpe. A pre-registered binary verdict set anywhere inside that band is a coin flip dressed as a gate, and the coarsening argument from the parent brief does nothing about it, because coarsening limits what the holdout leaks, not what it can resolve. One further asymmetry favours false passes: our Monte Carlo shows the naive annualized Sharpe carrying a finite-sample upward bias of +0.26 at T=300 and +0.16 at T=1,000 under SPY-like dynamics, which the power calculation above does not account for.
Three corrections to what the pipeline was likely to build, one of them a reversal of our own first pass. First, the robust time-aggregated estimator at q=252 is close to a wash, not clearly negative, and should be settled by Lo's pre-test rather than by the efficiency arithmetic. Lo's Eq 27 evaluated at SPY's measured Sharpe of 0.548 gives a variance ratio of 1.099, a 4.9 percent wider standard error, against a 7.3 percent narrower standard error from the rho of -0.076. We initially read this as net negative by evaluating Eq 27 at an assumed Sharpe of 1.0 (where the cost is 15 percent); at the measured Sharpe the comparison flips. The reason not to use it is simpler and is Lo's own: Ljung-Box does not reject on the ex-COVID series, so there is no established autocorrelation to correct. Note also that the 7.9 percent eta(252) figure is a point-estimate bias, a different quantity from a standard-error gain, and the two should not be netted against each other. Second, do not use a Mertens or Probabilistic-Sharpe plug-in standard error here. It buys 2 percent at daily frequency, it requires a fourth moment our own fitted parameters say is infinite, and its kurtosis input moves by a factor of 2.7 on the real data depending on one 50-day window. Use the Ledoit-Wolf studentized circular-block bootstrap, which is studentized rather than plug-in; their paper hands over the two-sample case and leaves the single-Sharpe extension to the reader. Third, retire L_min as a queryability threshold rather than raising it, because no reachable value makes the query informative, and replace it with a pre-registered power requirement computed per query.
What to build instead, and the tension in it. The paired difference against buy-and-hold is the only route within an order of magnitude of our data, and wet-rehearsal Bar 3 is already specified that way, so this is a cheap amendment rather than a redesign. Testing a candidate against the benchmark on the same bars collapses the standard error through the cross-correlation: computed by us from the Jobson-Korkie/Memmel variance on a consistent closed-form iid basis (7,927 bars for the single-Sharpe case, against the 7,985 the simulated iid control gives), L_min for a 0.50 gap falls to 1,589 bars at a return correlation of 0.90 and 797 at 0.95. Our Monte Carlo confirms the benefit survives GARCH but is materially eroded by it, giving roughly 2,450-2,770 bars at correlation 0.90 and 1,615-1,846 bars at 0.95. Under realistic dynamics no tested correlation fits inside the 1,500-bar cache; the 0.95 case needs 6.4 to 7.3 years against the 5.98 available, so this route is close but not yet reachable, and it becomes reachable with roughly one more year of accrual. The tension to name rather than hide is that correlation and edge trade off: a candidate 95 percent correlated with SPY is very nearly SPY, so the regime where the test has power is the regime where the difference being tested is hardest to generate. The amendment to Bar 3 is therefore to pre-register the expected candidate-benchmark correlation alongside the Sharpe gap, compute the required bars from the pair, and refuse the holdout query when the pre-registered pair is unresolvable on the block, which makes the gate's own power an auditable pre-registered quantity. Cross-sectional sealing helps less than the parent hoped, and we can now supply the discount it asked for: using N_eff = N/(1+(N-1)*cbar), 50 symbols at average pairwise correlation 0.50 give N_eff = 1.96, 100 at 0.35 give 2.81, 100 at 0.20 give 4.81. On the fitted-spec curve those reduce the 43-year requirement for a 0.50 gap to 20.4, 13.7 and 7.5 years, none reachable. Making a 0.50 gap resolvable on the existing 1,500 bars would need N_eff of 5.92, which requires average pairwise correlation below 0.17 across a large symbol set. Equities do not offer that within a regime. Breadth is a genuine 2-to-5x multiplier worth building, not a solution.
Why this is in the vault
This closes open-follow-up #4 of [[2026-08-31-sealed-holdout-refresh-policy]] and changes three specific build inputs for validation/holdout.py in /Users/ray/Projects/automated-investing: L_min = 1,000 daily bars is retired as a queryability threshold in favour of a per-query power requirement; wet-rehearsal Bar 3's undefined "materially higher Sharpe" gains a pre-registered Sharpe gap plus an expected candidate-benchmark correlation, with the query refused when the pair is unresolvable on the block; and the standard error is computed by studentized block bootstrap rather than a Mertens or Probabilistic-Sharpe plug-in. It also supplies the effective-independence discount that brief left unspecified for cross-sectional sealing (N_eff caps at roughly 2 to 5, needing 5.92 to close the gap), and it rules out Lo's robust q-aggregated estimator at q=252 on Lo's own Ljung-Box grounds rather than on the efficiency argument our first pass used.
Open follow-ups
- Derive the single-Sharpe studentized circular-block bootstrap that Ledoit & Wolf explicitly leave to the reader (p.858), and calibrate the block length against the GARCH persistence of 0.96 to 0.97 fitted here. It is the only standard error on the table valid under both fat tails and variance dependence, and nobody has written down the one-sample version we would actually call.
- Establish whether the Deflated Sharpe Ratio survives a non-existent fourth moment. Its denominator is the Mertens variance, so the infinite-fourth-moment result on our own fitted parameters propagates into every DSR the pipeline would report, and [[2026-06-01-cpcv-deflated-sharpe-autoinv-validation]] adopted DSR without this caveat.
- Resolve the Bailey-Lopez de Prado versus Ledoit-Wolf disagreement over whether Opdyke's (2007) stationary-case variance generalises Mertens or is simply wrong. The paper is paywalled at Springer; the disagreement is load-bearing because it determines whether the DSR expression is valid under serial dependence at all.
- Re-estimate the convergence-rate exponent on real multi-decade data rather than simulation. The T^-0.448 to T^-0.481 exponents behind the headline are Monte Carlo on GARCH processes fitted to a single 1,508-bar window, and the parameter fit is itself noisy; the same exercise on 70 years of actual S&P 500 daily returns would confirm or correct the number the L_min multiple is most sensitive to, and would test whether the exponent is stable across regimes. The sensitivity spread already seen here (0.378 at persistence 0.99 against 0.481 on the ex-COVID fit) is wide enough to matter.
- Determine what gate the holdout can decide at 1,000 to 1,500 bars with adequate power. Maximum drawdown, hit rate, turnover-adjusted net return and tail-loss bounds have different noise properties, and Bar 3 already pairs Sharpe with a drawdown condition whose power nobody has computed.
- Quantify the attainable Sharpe improvement at candidate-benchmark correlation near 0.95, so the pre-registered pair in the amended Bar 3 is feasible rather than aspirational. If no real overlay can beat SPY by 0.5 while staying 95 percent correlated to it, the paired-difference route has a ceiling we should know before building to it.
- Correct the finite-sample upward bias in the naive annualized Sharpe (+0.26 at T=300, +0.16 at T=1,000 measured here). The bias favours false passes at exactly the sample lengths the pipeline will use, and Christie (2005) is cited in the literature as having a small-sample adjustment we have not read.
Related
- [[2026-08-31-sealed-holdout-refresh-policy]] — the parent brief whose open-follow-up #4 this closes; source of the +/-0.92 figure, the 1,000-bar L_min and the reseal-with-budget policy
- [[2026-05-29-markov-system-requirements-v0]] — the wet-rehearsal plan whose Bar 3 is the actual holdout query; its undefined "materially higher Sharpe" is what this brief quantifies
- [[2026-06-01-cpcv-deflated-sharpe-autoinv-validation]] — DSR and MinTRL mechanics; its "T = return observations" caveat and 30-observation floor are the constraint this brief quantifies
- [[2026-07-05-autoinv-cpcv-vs-walk-forward-confirm-stage]] — the two-CV-regime confirm stage that must carry the inference load the holdout cannot
- [[2026-06-26-strategy-discovery-loop-architecture]] — the five-stage loop whose single sealed-holdout arbiter this brief constrains
- [[2026-05-29-strategy-pipeline-architecture-v0]] — the v0 architecture specifying the sealed holdout and the Phase-2
validation/holdout.pybuild this changes - [[2026-05-31-ensembles-systematic-trading-overfitting]] — the gates-not-reports posture that motivated a binary holdout verdict, necessary but shown here to be insufficient
- [[2026-06-01-ensemble-admission-correlation-threshold]] — ensembles are re-scored on the holdout as their own candidates, so they inherit this power limit
Sources
Vault:
~/rdco-vault/06-reference/research/2026-08-31-sealed-holdout-refresh-policy.md— parent brief; +/-0.92, L_min = 1,000 bars, reseal-with-budget, open-follow-up #4, the "optimistic floor" caveat at line 42~/rdco-vault/01-projects/investing/2026-05-29-markov-system-requirements-v0.md— line 123, wet-rehearsal Bar 3, the benchmark-relative holdout query~/rdco-vault/06-reference/research/2026-06-01-cpcv-deflated-sharpe-autoinv-validation.md— DSR / MinTRL formulas and the 30-observation floor~/rdco-vault/06-reference/research/2026-07-05-autoinv-cpcv-vs-walk-forward-confirm-stage.md— two-regime confirm stage decision~/rdco-vault/06-reference/research/2026-06-26-strategy-discovery-loop-architecture.md— five-stage discovery loop~/rdco-vault/01-projects/investing/2026-05-29-strategy-pipeline-architecture-v0.md— sealed-holdout spec and Phase-2 build plan~/rdco-vault/06-reference/research/2026-05-31-ensembles-systematic-trading-overfitting.md— gates-not-reports posture~/rdco-vault/06-reference/research/2026-06-01-ensemble-admission-correlation-threshold.md— ensemble admission and holdout re-scoring
Repo and data (verified 2026-10-29):
/Users/ray/Projects/automated-investing/data_cache/SPY_2018-01-01_2023-12-31_5f6c0c96.csv— 1,509 price rows, 1,508 log returns, 5.98 years. Measured by us: annualized Sharpe 0.548; g3 = -0.783 and raw kurtosis 14.885 full window; g3 = -0.476 and raw kurtosis 5.581 excluding 2020-02-20 to 2020-04-30; rho_1 = -0.145 full and -0.020 ex-COVID; Ljung-Box(5) on returns 55.5 full / 2.99 ex-COVID; Ljung-Box(5) on squared returns 1,311 full / 118.6 ex-COVID; GARCH(1,1) Gaussian quasi-maximum-likelihood with variance targeting alpha=0.192/beta=0.778 full and alpha=0.148/beta=0.814 ex-COVID, standardized-residual kurtosis 4.82 and 4.90
Computed by us (scripts and raw output in /Users/ray/scratch/sharpe-se/, seeded and reproducible):
part1_analytic.py— reproduction of the parent brief's iid figures; AR(1) and Mertens standard-error grids over SR in {0, 0.5, 1.0}, T in {300, 500, 1000, 2000, 2520}, rho in {-0.20, -0.10, -0.05, 0, 0.05, 0.10, 0.20}; Lo Eq 22 eta(252) table; Lo Eq 27 efficiency cost; L_min tables; Jobson-Korkie/Memmel paired-difference tablepart2_montecarlo.py— 40,000-path validation of the AR(1) derivation (agreement within 1 percent); finite-sample Sharpe bias; first detection of the growing GARCH errorpart3_garch.py— GARCH(1,1) fourth-moment condition; 30,000-path convergence tests to T=10,080; decomposition showing the inflation is entirely in the variance estimatorpart4_real_and_lmin.py— real SPY moment and Ljung-Box diagnostics; fitted standard-error scaling exponents; L_min solved from the fitted curves; paired-difference test under GARCHpart5_summary.py— headline L_min multiples and resolvable-gap figurespart6_fixes.py— GARCH(1,1) fit to the cached SPY series; decomposition of the gap into iid and non-iid legs; cross-sectional breadth on the fitted curve; Lo Eq 27 at the measured Sharpe; local scaling exponent and extrapolation sizingpart7_fitted_specs.py— the headline run. Monte Carlo at the four GARCH specs actually fitted to the cached SPY series (full-window and ex-COVID, each with the measured and the literature rho), 20,000 replications per cell over T in {300, 500, 1000, 2000, 2520, 5040, 10080}, plus iid and higher-persistence controls. This supersedes the assumed-parameter specs used in parts 2-4 for the headline L_min; the control run reproduces the earlier assumed-spec figure (12,967 against 13,123) confirming the two harnesses agree
Web:
- Andrew W. Lo — "The Statistics of Sharpe Ratios," Financial Analysts Journal 58(4), July/August 2002, pp. 36-52 — https://traders.studentorg.berkeley.edu/papers/The-Statistics-of-Sharpe-Ratios.pdf (PRIMARY REACHED; published version with the AIMR 2002 footer. Eq 8/9 iid SE, Eq 14/15 GMM, Eq 18 annualization, Eq 20/22 eta(q), Eq 25-27 and A32/A33 robust-estimator variance, Eq 28 Ljung-Box, Tables 2-4. Lo's own page at alo.mit.edu 404s for this paper. Two printing errata relevant to anyone re-deriving: body Eq 25 omits the leading q that appendix Eq A32 carries, and the Newey-West truncation condition on p.48 is printed inverted)
- Baltussen, van Bekkum & Da — "Indexing and stock market serial dependence around the world," Journal of Financial Economics 132(1), 2019, pp. 26-48 — https://www3.nd.edu/~zda/Indexing.pdf (PRIMARY REACHED; Table 1 S&P 500 daily AR(1) +0.103 pre-1999 / -0.076 post-1999, and the one-day-implementation-lag caveat)
- Ledoit & Wolf — "Robust performance hypothesis testing with the Sharpe ratio," Journal of Empirical Finance 15 (2008) 850-859 — https://uzh.ch/cmsssl/econ/dam/jcr:ffffffff-935a-b0d6-0000-00007214c2bc/jef_2008pdf.pdf (PRIMARY REACHED; studentized circular-block bootstrap for the Sharpe difference, Remark 3.1 taxonomy of Lo 2002, rejection of Opdyke's stationary-case formulae, single-Sharpe extension left to the reader on p.858)
- Bailey & Lopez de Prado — "The Sharpe Ratio Efficient Frontier," Journal of Risk 15(2) — https://www.davidhbailey.com/dhbpapers/sharpe-frontier.pdf (PRIMARY REACHED; Eq 8 Mertens form, the collapsed (g4-1)/4 form, Probabilistic Sharpe Ratio Eq 11, the Christie 2005 / Opdyke 2007 equivalence claim, and the recommendation to substitute bounds for unstable g3/g4 estimates)
- Bailey & Lopez de Prado — "The Deflated Sharpe Ratio" — https://www.davidhbailey.com/dhbpapers/deflated-sharpe.pdf (PRIMARY REACHED; DSR Eq 2 using the same Mertens denominator)
- Deng, Dulaney & McCann (SLCG) — "Sharpe Ratio Practice Note," 2012 — https://www.slcg.com/files/practice-notes/sharpe-ratio.pdf (PRIMARY REACHED; independent statement of the Mertens variance with explicit raw-kurtosis convention, plus a worked MinTRL example)
- Elmar Mertens — "Comments on variance of the IID estimator in Lo," working paper, University of Basel, 2002 — UNREACHABLE; absent from the author's site, 404 at the circulated URL, no Wayback snapshots. The formula is UNVERIFIED-TO-PRIMARY, cited via the two independent secondaries above which agree exactly, and independently recoverable from Lo Eq A32 at q=1
- J.D. Opdyke — "Comparing Sharpe ratios: So where are the p-values?", Journal of Asset Management 8(5), 2007, pp. 308-336 — https://ideas.repec.org/a/pal/assmgt/v8y2007i5d10.1057_palgrave.jam.2250084.html (ABSTRACT ONLY; body PAYWALLED at Springer, noted and not retried. His Eq 3 and Eq 6 were not read, so the Bailey-Lopez de Prado versus Ledoit-Wolf disagreement is unadjudicated here)
- Kim & White — "On more robust estimation of skewness and kurtosis," Finance Research Letters 1(1), 2004, pp. 56-73 — https://ideas.repec.org/a/eee/finlet/v1y2004i1p56-73.html (RECORD ONLY for the published version. The daily S&P 500 1982-2001 figures used above, skewness -2.39 / raw kurtosis 53.62 full and -0.26 / 6.80 excluding the 1987 crash, come from the 2003 working-paper version read on a third-party mirror and are UNVERIFIED-TO-PRIMARY)
- The "ordinary window" moment scenario (g3=-0.51, raw kurtosis 7.5) used in the Mertens grid derives from an unrefereed arXiv preprint (arXiv:2201.09434, Table IV, S&P 500 daily 2011-2016) and is treated here as ILLUSTRATIVE, not sourced; the "severe tails" scenario (g3=-0.5, raw kurtosis 11) is hypothetical. Neither is load-bearing, since the conclusion of that leg is that the correction is negligible at daily frequency either way, and the cached-SPY values measured directly above bracket both
- The GARCH(1,1) fourth-moment condition kalpha^2 + 2alpha*beta + beta^2 < 1 is standard in the GARCH literature and is attributed here with low confidence: Bollerslev (1986) is usually cited for the conditionally-Gaussian special case and the general kurtosis form appears in later work (He & Terasvirta 1999 is the usual reference). Not verified against either primary. The condition is applied only to classify our own fitted parameters, and the qualitative finding is corroborated independently by the simulation convergence behaviour in
part3_garch.py