Strateva

Audited product boundary

What the engine does — and where it stops.

This is the operational boundary of the service available today, checked against the platform and quantitative packages on 22 August 2026. A capability is not presented as included when its data, protocol or evidence requirements are not met.

What sponsored validation includes

One automated validation: one exact strategy version × one tradable asset × one timeframe × one confirmed historical period.

  • Intake through the Strateva MCP from Claude or ChatGPT, using code, Pine Script, a trade list or precise written rules.
  • AI-assisted normalization and preparation of a separate execution adapter. The original source remains immutable and hashed.
  • Frozen scope, data identity, costs, benchmark and decision criteria before execution.
  • Isolated execution without network access, followed by a verified metric-free StrategyTrace.
  • Closed-engine metrics, classic moving-block Monte Carlo, Fisher–Yates permuted Monte Carlo, deterministic gates and a PDF report.
  • An unfavourable or inconclusive result is a completed delivery. Access and optional support never depend on the verdict.

Automation, not an analyst engagement

The sponsored service does not include routine manual review by a quantitative analyst. Contract checks, hashes, execution guards, evidence reconciliation and verdict construction are automated. AI prepares material and explains existing results; it cannot calculate official metrics, enter the kernel or choose the verdict.

Three evidence levels

Know what kind of evidence you are buying.

Every report names the highest level actually reached. Higher levels require the protocol and time they claim; they are never inferred from a strong backtest.

  1. Included now

    1. Reproduced history

    The frozen strategy is replayed on the confirmed historical data, with declared costs, a benchmark when valid, robustness simulations and a reproducible evidence chain. This is historical evidence, not out-of-sample proof.

  2. When eligible

    2. Walk-forward OOS

    Chronological folds test decisions on later historical observations. Rolling WFA evaluates one frozen candidate; nested WFA is used only when multiple candidates were frozen before selection. It is stronger historical evidence, but still historical.

  3. Requires future data

    3. Prospective holdout

    The strategy, costs, benchmark, horizon and decision rule are sealed before new observations exist. Results remain hidden until the horizon completes. It cannot be delivered immediately or reconstructed from old data.

These are evidence levels within one service, not performance ratings. A result may be unfavourable or not evaluable at any level.

End-to-end path available today

  • The automated beta data connector currently acquires daily data only.
  • Its current acquisition contract accepts cash equities/ETFs and spot FX through the supported Yahoo path, plus supported spot-crypto pairs through Binance. A request that cannot be acquired must stop before execution.
  • Intraday bars exist in wider contracts but are not advertised as an automated end-to-end data service today.
  • The current validation scope is one tradable asset. Internal multi-instrument schema capacity is not a portfolio service.
  • Derivatives are excluded: options, futures, perpetual swaps, CFDs, leveraged tokens and structured products.

Checks that apply only when the protocol supports them

  • Rolling walk-forward evaluates one frozen candidate through chronological out-of-sample folds when enough suitable observations exist. It performs no parameter selection.
  • Nested walk-forward applies only when at least two genuinely distinct candidates were frozen before execution. It requires at least two inner folds plus explicit purge, embargo, execution lag and selection rules.
  • A prospective holdout is not historical OOS. It begins after the strategy is frozen, accepts 2 to 5,000 expected observations, withholds metrics until the full horizon arrives, and therefore cannot be delivered immediately.
  • Benchmark metrics exist only when a valid benchmark series was frozen on the same calendar and period.
  • Monte Carlo results affect the verdict only through criteria registered before execution. They are not an automatic reason to pass or fail.
  • Deflated Sharpe Ratio (DSR) and PBO/CSCV are evaluated only when the author supplies the complete, synchronised family of variants tried and identifies the selected variant. With one winner or an incomplete family, the honest result is NOT EVALUABLE.

What the result never means

  • A Strateva report is an independent validation report, not a certification, accreditation, investment recommendation or guarantee that a strategy is safe to trade.
  • The controls can detect, reduce or disclose specific risks. They do not eliminate bias and do not prove that undiscovered bias is absent.
  • READY FOR SHADOWING means unfunded forward observation is the next research step. It is not permission to allocate capital.
  • No result guarantees profit, prevents loss or predicts future performance. Historical reproduction and historical OOS remain historical evidence.
  • DSR/PBO can adjust only for the complete trial family that was disclosed. Undisclosed variants remain an explicit blind spot.

Not included

  • No live trading, signals, brokerage connection, portfolio management or capital-allocation recommendation.
  • No guarantee of profit and no claim that historical robustness predicts future returns.
  • No routine human analyst review in the sponsored service.
  • No DSR/PBO conclusion when the complete set of tried variants and their synchronised return series are unavailable.
  • No automatic proof of market capacity, market impact, borrow availability, tax treatment or institutional point-in-time data quality.
  • No claim that walk-forward folds are independent experiments, or that nested walk-forward replaces a future prospective holdout.
Bias controls and their limits
  • The AI may repair only an unambiguous mechanical look-ahead issue by moving information or execution backward in time without changing the economic rule. Ambiguous cases must be rejected or clarified.
  • Static guards catch recognisable Python patterns such as negative shifts, centred windows, backward fills and same-close execution. A clean scan is triage, not proof that no look-ahead exists.
  • Static analysis cannot see vendor revisions, publication delays, SQL/API behaviour or data-vintage semantics. Those require point-in-time provenance and runtime assertions.
  • Survivorship bias is controlled only with point-in-time membership and delisted instruments for dynamic universes. Otherwise the PDF must state that survivorship bias is not controlled.
Audited numerical ceilings
  • StrategyTrace: up to 100,000 observations by contract; the deployed sandbox output channel is capped at 16 MiB, so bytes may bind first.
  • Monte Carlo: 100 to 100,000 repetitions, explicit seed and explicit block length. The client-facing summary is deliberately limited to P5, P50 and P95 instead of a dense table of deciles or percentiles.
  • Walk-forward protocol: at most 64 frozen candidates and 64 scalar parameters per candidate; these are safety ceilings, not a promise that every grid is statistically defensible.
  • Prospective holdout: 2 to 5,000 observations. The current internal SHA-256 chain detects rewriting inside the normal workflow but is not an independent public timestamp against an administrator controlling all hosts.
Version checked

StrategyTrace quantitative kernel 1.0.0 · selection-bias module 0.6.0 · validation-core 0.8.0. DSR/PBO still require a complete, synchronised variant matrix; one winning trace is not enough.

Have something that needs an independent check?

Describe the strategy, the backtest or the model through the MCP. You confirm the exact scope before anything runs.

Connect the MCP

Validation is sponsored. Optional support is separate and never affects the result.