VTR-1

Every track record in systematic finance is, today, graded by the party with the greatest incentive to shade it, and the tools that check it verify only that a number was reported, not that it is true. This standard exists to close that gap. It defines what it means for a performance claim to be verifiable: reproducible by an independent party, from committed inputs, with no trust placed in the issuer.

This standard governs the claim, not the strategy. A conforming credential proves properties of a track record while the strategy, its parameters, and (at Level III) its data remain private. It is asset-class agnostic: equities, futures, FX, commodities, crypto, and rates are all in scope where a canonical, committable price series exists.

A credential MUST declare exactly one Level. Higher Levels subsume the requirements of all lower Levels.

The claim: this performance actually happened, as stated, on real data, net of real costs.

A Level I credential MUST prove, in-circuit:

1. The reported return, Sharpe, drawdown, and worst-bar are the correct results of applying the committed strategy to the committed data.
2. The committed per-side costs were charged on every transaction; that they meet or exceed the floor is confirmed by the verifier (§4, step 5), with in-circuit floor enforcement RECOMMENDED.
3. No lookahead: the position at bar t depends only on data through bar t.
4. The annualization basis is consistent with the committed data window (calendar-honest; over-counting is rejected).
5. The strategy specification is absent from the proof, present only as a sealed commitment.

Level I answers did this happen? It does not address luck.

The claim: this result is not merely real, but unlikely to be luck given how it was found.

A Level II credential MUST, in addition to Level I:

1. Commit the complete trial ledger before evaluation, so that N is the leaf count of the committed tree and cannot be understated.
2. Force the presented winner, in-circuit, to be the maximum of the committed trials.
3. Recompute the anti-overfitting statistics inside the proof: at minimum the Deflated Sharpe Ratio (Bailey & López de Prado) using the committed N; a conforming implementation SHOULD also compute the Probability of Backtest Overfitting (CSCV) and MAY compute the Probabilistic Sharpe Ratio, the Superior Predictive Ability test, and combinatorial purged cross-validation over the committed set.
4. Report every computed statistic, pass or fail. A credential MUST NOT report a subset that flatters the result.

Level II answers is it skill or luck, given the search? It is the level at which the standard's central contribution, the committed trial ledger, binds.

The claim: everything above, provable by a manager who reveals nothing, to an allocator who trusts no one.

A Level III credential MUST, in addition to Level II:

1. Prove the track record while the strategy or model runs entirely on the submitter's own machine, submitting only committed decisions (the sealed path). The model MUST NOT be recoverable from the credential.
2. Bind an owner key so the credential's holder is provable.
3. Be extensible as a chained track: the same committed strategy re-proven over a growing data window, each link cryptographically bound to its parent, so that bars arriving after a commitment carry forward-evidence weight.

Level III is the allocatable credential: a manager proves an edge to an allocator without disclosure, and the allocator verifies it without trust.

A conforming verifier MUST, given a credential and the verifier's own copy of the data:

1. Confirm the proof's guest image hash equals the era the credential names.
2. Verify the cryptographic proof.
3. Re-derive the data root from the verifier's own data and confirm it matches the committed root.
4. Confirm the gate policy committed in the proof is the policy the implementation publishes for that era and Level (§2), and that it was not selected per credential.
5. Confirm costs meet or exceed the floor and the annualization basis is bound to the data.
6. Confirm the trial count is the committed leaf count (Level II+) and the winner is its maximum.
7. Emit the verdict. The verifier MUST return a failure indication on any check that does not pass, and MUST NOT require trust in the issuer at any step.

A verification implementation SHOULD be distributable and runnable offline. The reference verifier completes in milliseconds and is available under an open license.

Stated normatively, because the boundary is the standard's integrity:

• A credential is a seal on history, not a forecast. It MUST NOT be represented as predicting future returns.
• It does not model capacity, market impact, or fill quality beyond the committed cost floor, and MUST NOT be represented as doing so.
• Fills that the committed data cannot justify are out of scope. Bar data cannot establish that a trade occurred at a price the bar's range does not contain; a proof of a fill that did not happen is a proof of fiction. A conforming implementation MUST either refuse such a strategy, or evaluate it under the conservative fill discipline of R-27 and disclose the assumptions made. Path-dependent exit logic (profit targets, stop levels, time-based exits) is permitted on that condition and on no other.
• High-frequency strategies are out of scope permanently, without exception, because their economics live below the resolution of any committable bar series. No fill discipline rescues them, and a conforming implementation MUST refuse them.
• The trial ledger binds the committed search. Trials run off-ledger are not covered; closing that residual requires forward pre-registration (an anchored ledger timestamped before the out-of-sample bars exist). A credential MUST disclose whether it is a sealed backtest or a pre-registered forward record.

The engine that issues credentials evolves; the standard governs that evolution so that no issuer, including the standard's author, can retroactively change what a past credential means.

1. A credential names its judge. It commits the image hash of the era that produced it.
2. Eras are superseded, never revoked. A changed engine is a new era with a new hash; old credentials verify against their own era permanently.
3. A changed judge is a new judge. Any change reaching the compiled proving program is, by definition, a new era.
4. Weakness is disclosed, not rewritten. A soundness gap in a frozen era is published as status, with the credential remaining checkable; the reader, not the issuer, decides its worth.

This specification is itself versioned under the same discipline. This is Version 1.1 (editorial revision of the frozen 1.0 normative text — 0 requirement changes); future versions supersede but never invalidate credentials issued under prior versions.

Anyone MAY implement this standard. A party claiming conformance MUST publish, for each credential: its declared Level, the era hash, the committed data root, and a re-verifiable proof object. A party MUST NOT describe a credential as conforming to a Level whose requirements it does not meet.

A reference implementation, a public registry of conforming credentials, including the issuer's own published failures, and an open verifier are maintained at mizan.market. Conformance vectors for testing an independent implementation are published under Appendix C. This specification makes no claim about the relative merit or priority of any implementation, including its steward's. Conformance is established by the requirements of Appendix B and the vectors of Appendix C. It is tested, not asserted.

This standard implements published statistical and cryptographic work; the full citation lineage is maintained at mizan.market/research/lineage. None of the cited authors has endorsed this standard; the citations run one direction. This document implements their work, and its author takes responsibility for the implementation.

Corrections to this specification are invited and will be published, with attribution, under the same discipline applied to the reference implementation's own numbers: [email protected].

Errata. The normative text being frozen, a correction to it is published as an erratum rather than applied silently to the document. An erratum records what was found, who found it, and what follows: no change where the text is defensible as written, an entry in the errata list where the text is unclear, or a numbered successor version where a requirement genuinely changes. Errata are published whether or not the finder is named, and whether or not the finding flatters the standard. Send them to [email protected].

Reporting a soundness break. A conforming implementation MUST publish a route by which a suspected soundness break can be reported, and MUST disclose a confirmed break as era status under §6. The affected era is marked, not rewritten, and its credentials remain checkable so that readers can judge them for themselves. For this specification and its reference implementation that route is [email protected]. A report will be acknowledged, and a confirmed break published, whether or not the finder wishes to be named. An implementation that quietly patches a soundness break has revoked a credential without saying so, which §6 forbids.

A conforming implementation MUST disclose, on every credential, which representation the evaluated strategy took. Two are defined.

Declared representation. The strategy is expressed in a specification language the implementation evaluates directly. The representation is legible to anyone holding the specification, but the implementation's language bounds what can be expressed. An implementation MUST NOT silently substitute a strategy it can express for one it cannot: if a submitted strategy is not expressible, the implementation MUST refuse, or disclose the substitution on the credential.

Committed representation. The strategy is executed by its owner, which commits only its resulting decisions, and the implementation evaluates the committed decision stream. Any strategy is expressible in this form, including path-dependent logic such as profit targets, stop levels and time-based exits. The logic itself is not revealed and is therefore not legible from the credential alone.

The two are complementary rather than ranked. A submitter may obtain legibility and expressiveness by publishing a deterministic encoder alongside a committed-representation credential, so that a reader verifies strategy-to-decisions with the encoder and decisions-to-metrics with the proof. This specification does not require any implementation to support arbitrary strategy expressiveness in its declared representation, and an implementation MUST NOT imply that it does.

This section states the adversary model and the attack classes a conforming implementation is expected to address. It is stated normatively because a verification standard that does not name its attacks is decoration.

Adversary model. The adversary is the submitter, and in the limit the issuer. A conforming implementation MUST be sound against a submitter who controls their own inputs and is motivated to obtain a favourable verdict, and its governance (§6) MUST be sound against an issuer motivated to change what a past credential meant.

Trial-set padding. Duplicate or near-duplicate candidates inflate the trial count while adding no dispersion, which can collapse a selection-bias correction toward neutrality. A conforming implementation SHOULD refuse identical and exactly proportional candidates, and MUST disclose any residual it does not detect.

Decoy candidates. Deliberately degenerate entries can flatter a favoured result by comparison. Corrections robust to this class, such as studentized multiple-testing procedures, SHOULD be preferred over those that are not.

Annualization inflation. Overstating periods per year inflates every annualized statistic. An implementation MUST bound the declared periodicity against the committed data window.

Data substitution. Evaluating on data other than that claimed. Prevented by requiring the verifier to re-derive the data commitment from its own copy.

Off-ledger search. Candidates evaluated outside the commitment are invisible to the correction. This is a disclosed residual, not a solved problem; closing it requires forward pre-registration.

Counterfactual fills. Path-dependent exit logic can book prices the market never traded, for example an exit at a stop level the instrument gapped past. An implementation evaluating such logic MUST use conservative fill assumptions, MUST disclose them, and MUST NOT assume a fill it cannot justify from the committed data. A proof of a fill that did not happen is a proof of fiction, and encoding one inside a credential is worse than not offering the feature.

Issuer capture. Retroactive alteration of the judging program. Prevented by era law (§6): the credential commits the image of its judge, programs are superseded, never revoked, and archived builds remain independently runnable.

This standard specifies enforcement for statistical corrections it does not originate. The deflation program derives from Bailey and López de Prado, including the Probabilistic Sharpe Ratio (2012), the Deflated Sharpe Ratio (2014), and the Probability of Backtest Overfitting via combinatorially symmetric cross-validation (Bailey, Borwein, López de Prado and Zhu, 2017), together with purged and embargoed combinatorial cross-validation (López de Prado, 2018). The multiple-testing program derives from White's Reality Check (2000), Hansen's test for Superior Predictive Ability (2005), and the multiple-testing haircuts and minimum backtest length of Harvey and Liu (2015). The characterisation of why a researcher cannot reliably count their own trials after the fact is due to Gelman and Loken (2014). The cryptographic constructions rest on Goldwasser, Micali and Rackoff (1985, 1989), Merkle (1987), and the transparent proof systems of Ben-Sasson and collaborators (2018). Full citations are maintained at mizan.market/research/lineage.

None of the cited authors has reviewed or endorsed this specification.

An implementation claiming conformance SHOULD be able to answer each of the following affirmatively, and a party evaluating a conformance claim MAY use this list directly.

1. Does every credential name the exact judging program by cryptographic image, and does an archived build of that program still verify it?
2. Can a verifier re-derive the data commitment from a copy of the data obtained independently of the issuer?
3. Is the trial count derived from a commitment made before evaluation, rather than declared?
4. Is the presented winner constrained to be the maximum of the committed set?
5. Are all computed statistics reported, including those that fail?
6. Are costs charged on every transaction, and is the floor checked at verification?
7. Is the annualization basis bounded by the committed data window?
8. Is the strategy representation (§9) disclosed on the credential?
9. Are FAIL credentials issued, published, and verifiable on the same terms as PASS credentials?
10. Are known unaddressed attack classes disclosed rather than omitted?

An implementation that answers ten of ten and publishes its refusals is conforming in substance as well as in form. An implementation that publishes only its passes has not been evaluated by anyone, including itself.

Every normative requirement in this specification is assigned a stable identifier below. A conformance claim SHOULD reference these identifiers, so that a claim is precise and a gap is nameable. "Conforms to VTR-1 Level II except R-27" is a meaningful statement; "conforms to VTR-1" alone is not. Identifiers are stable across revisions of this version: a requirement that is removed retains a retired identifier rather than having its number reused.

Known-answer tests for this specification are published as VTR-1 Conformance Vectors v1.0, a self-contained offline kit of nine vectors: five that MUST verify, three that MUST NOT, and one published honest failure. Each vector names the requirement identifiers it exercises, and each states a normative exit code, so that conformance can be tested automatically rather than asserted.

The negative vectors carry the most weight. One presents a byte-mutated copy of the committed data and MUST be rejected (R-01). One presents a sealed credential without its aggregate and MUST fail closed rather than degrade to a partial pass (R-11). One presents a credential to the engine of a different era and MUST fail on the image identifier — the governance model of §6, observable as a test.

The kit is published under CC BY 4.0 and exists for a single purpose: to make an independent second implementation cheaper to build. A specification with one implementation is a description of that implementation. The author cannot close that gap alone; publishing the tests is the part of it that is within his control.

Download the conformance vectors →

Sections 3 and 4 require an implementation to recompute the anti-overfitting statistics. Those requirements are not reproducible unless the estimators and the arithmetic are pinned, because the published literature admits several conventions and two honest implementations choosing differently will disagree on identical inputs. This appendix pins them. An implementation that follows the cited papers but not this appendix may be statistically defensible and still non-conforming, because it will not reproduce the conformance vectors of Appendix C.

D.1 Arithmetic

R-28 (MUST). All statistics that a credential commits or a verifier re-derives are computed in deterministic fixed-point integer arithmetic with scale S = 10⁶; a value v is represented by the integer round(v · S). Floating-point arithmetic MUST NOT be used anywhere in the evaluated path. Division truncates toward zero. Intermediate products are computed at 128-bit width before rescaling.

R-29 (MUST). Verification equality is exact. There is no numerical tolerance, and a conforming implementation MUST NOT introduce one. Because the arithmetic is integer and deterministic, two conforming implementations agree bit-for-bit or one of them is wrong. This is a deliberate departure from ordinary quantitative practice, where a tolerance would be normal: a verifier that accepts "close enough" cannot distinguish a rounding difference from a manipulated statistic.

D.2 Moments of the return series

R-30 (MUST). Given n per-period returns r_i, an implementation computes population moments (denominator n, not n−1), standardising deviations before taking third and fourth powers:

The kurtosis convention is the most common source of silent disagreement between implementations, and the per-period Sharpe convention the second. Both are stated here for that reason. An implementation returns "undefined" rather than a value for n < 2 or zero variance.

D.3 Probabilistic Sharpe Ratio

R-31 (MUST). The PSR is the Bailey & López de Prado estimator, evaluated on the moments of D.2:

The radicand is floored at the smallest positive representable value, so that a pathological moment combination degrades the statistic rather than dividing by zero.

D.4 The deflation benchmark

R-32 (MUST). The benchmark SR₀ is the expected maximum Sharpe of N independent trials, with γ the Euler–Mascheroni constant (0.5772156…) and σ_SR the population standard deviation (denominator N) of the per-period Sharpe across the resolved trials of the committed ledger (unresolved entries are excluded from the dispersion but still counted in N):

For large N the quantity 1/N falls below the representable grid and forming (1 − 1/N) loses it entirely. An implementation MUST therefore evaluate the upper quantile directly from −ln(1/N) = ln N in that regime rather than through the complement. An implementation that does not will silently saturate SR₀ as the trial count grows, producing a benchmark that stops rising exactly when the search gets large enough to matter. That is the failure mode the deflation exists to prevent.

D.5 Serial correlation

R-33 (MUST). The sample length entering the PSR is the effective sample size under AR(1) serial correlation, with ρ the lag-1 autocorrelation of the return series:

The correction is deliberately one-sided. Positive autocorrelation shrinks the effective sample, widening the sampling distribution and making the significance test harder. Negative autocorrelation is floored at zero rather than being allowed to inflate n, so the correction can never flatter a result. n_eff never exceeds n.

D.6 Deflated Sharpe Ratio and Minimum Backtest Length

R-34 (MUST). DSR = PSR(SR₀) evaluated with n_eff in place of n, using the moments of D.2, the benchmark of D.4 and the correction of D.5. The Minimum Backtest Length is the smallest sample length at which the observed per-period Sharpe exceeds the noise ceiling of N trials, n > (E_unit/ŜR)², rounded up and floored at 2, against the claimant in both directions.

D.7 Annualisation

R-35 (MUST). Where an annualised figure is reported, the factor is √(periods per year) applied to the per-period Sharpe, and the declared periods-per-year MUST be bounded against the committed data window per R-26. All statistical tests in this appendix are performed on per-period quantities; annualisation is a reporting transformation applied afterwards and MUST NOT enter any significance calculation.

D.8 Probability of Backtest Overfitting (CSCV)

R-39 (MUST where computed). PBO is computed by combinatorially symmetric cross-validation (Bailey, Borwein, López de Prado and Zhu). The T observations are partitioned into S disjoint, contiguous, equal-length sub-periods. For every balanced partition of those sub-periods into a training half and its complement, all C(S, S/2) of them with each partition and its complement both counted, the trials are ranked in-sample, the in-sample best is selected, and the procedure records whether that trial's out-of-sample rank falls below the median. PBO is the fraction of partitions for which it does.

Three properties are normative. The split is canonical, not prover-chosen. An implementation that lets the submitter influence the sub-period boundaries has not computed PBO; it has computed a number the submitter selected. Every trial's per-sub-period Sharpe is re-derived from its committed specification on the committed bars, never declared. The procedure fails closed: any binding failure or degeneracy yields PBO = 1.0, which fails any sane threshold, rather than an error the reader might overlook.

PBO measures overfitting of the selection among the committed trials. It does not establish that the committed set was the whole search; that residual is §10, off-ledger search.

R-01 requires a verifier to re-derive the data root from its own copy of the data, and §2 derives the trial count from the leaf count of a commitment. Neither is reproducible unless the constructions are pinned. This appendix pins them. Appendix D makes the statistics reproducible; this appendix makes the commitments reproducible. An implementation needs both to pass the vectors of Appendix C.

E.1 Hash and domain separation

R-40 (MUST). All commitments use SHA-256. Every hash input begins with a domain tag or a single-byte prefix, so that a value hashed in one role can never be reinterpreted in another. The tree prefixes are:

Named domain tags are used where a structure is not a bar tree, for example the trial-ledger leaf domain MIZAN_TRIAL_LEAF_V2. A tag is versioned, and a changed encoding takes a new tag rather than reusing the old one, so a commitment made under one encoding can never validate under another.

E.2 Data leaves

R-40 (MUST). A data leaf binds the bar's position and its full OHLC, each as little-endian bytes, in this order:

Binding the index means bars cannot be reordered without changing the root. Binding high and low means intrabar extremes are tamper-bound exactly as closes are, so a prover cannot shave a wick to soften a drawdown that the gate marks against. An implementation that commits only closes is not conforming, and the difference is not cosmetic: it is the difference between a drawdown that is checked and one that is asserted.

E.3 Tree construction

R-40 (MUST). Internal nodes are SHA-256( 0x01 ‖ left ‖ right ). Where a level has an odd number of nodes, the unpaired node is promoted unchanged to the next level; it MUST NOT be hashed against a copy of itself. Duplicating the final node admits distinct leaf sets that produce identical roots. This is the malleability class recorded as CVE-2012-2459, and an implementation that duplicates is unsound regardless of what else it does correctly. The empty tree hashes to SHA-256( 0x02 ‖ MIZAN_EMPTY ).

E.4 Trial-ledger leaves

R-40 (MUST). A trial leaf binds its sequence position, the commitment to the trial's specification, its registration time, and its resolved result:

Because seq and the result are both bound, trials can be neither reordered nor re-scored after commitment without changing the root. That is the property that makes N structural instead of reported. An unresolved trial carries a reserved sentinel in the result field; it counts toward N and is excluded from the dispersion of R-38.

Version 1.1 is the current rendering — an editorial revision of the frozen 1.0 normative text, with 0 requirement changes. The frozen normative text has not changed and will not change; the mathematics of Appendix D was re-set as display typography on 1 August 2026, and each superseding rendering supersedes the one before it. All are published, all are timestamped, and the older ones are retained rather than withdrawn — the same superseded-never-revoked rule §6 applies to engine eras.

A copy whose digest appears in none of these lists is not this document. Verify with shasum -a 256 and check the accompanying .ots proof. The renderings carry byte-identical normative text — all 85 normative sentences and all 41 requirement identifiers are verbatim the same, which is machine-checkable by extracting the text of each. Only the typesetting of the equations differs.

All renderings, conformance vectors and timestamps →

VTR-1 is an open specification. Anyone may implement, cite, or extend it with attribution under CC BY 4.0, including parties who compete with its steward. It is deliberately not named after the organisation that authored it: a standard carrying its issuer's name in its title cannot be adopted by that issuer's competitors, and a standard nobody else can adopt is not a standard. MIZAN stewards VTR-1; it does not own the conformance of others. Implementations are invited, and an independent implementation that reproduces the conformance vectors of Appendix C will be listed here on request.