Reconstructing an AI signal record
Versioned outputs and explicit evaluation conventions before performance claims.
Prediction sequence
Request consecutive outputs with timestamps, model versions and the definition of the target or decision. Include missing or invalid outputs. A selected list of winners cannot supply the denominator.
Evaluation layers
Recompute the stated prediction metric separately from a simulated trade return. Preserve the chosen research-fill convention and costs; do not substitute a more favourable one. Actual subscriber executions require their own evidence.
Aggregation
State the period, exposure and formula for any combined figure. Summing model percentages does not automatically create portfolio growth. If a rounded summary cannot be reconciled from records, report the gap instead of inventing a compounding explanation.
Sources
- QuantConnect optimization parameters
- ALFRED real-time and vintage data
- OpenTimestamps protocol documentation
Research guides
AI data leakage and holdout audit
Audit feature availability, outcome overlap and test exposure with a frozen experiment record and a worked calendar-boundary example.
Backtest, forward and live evidence
Reconcile simulated decisions, prospective instructions and account fills without promoting one evidence layer into another.
Algorithmic versus AI signal claims
Map fixed rules, trained predictions, AI explanations and order routing to the specific consumer evidence each needs.
Reproduction needs the original convention
An invented report retains data and a model but changes the hypothetical entry window after inspecting outcomes. Its new result is a different research object, not a byte-faithful reproduction of the original. Preserve the original best hypothetical research-fill policy and its frozen window where supplied, while keeping actual subscriber executions separate.
A complete run manifest also retains target, transforms, package versions and population. If data or calibration support are missing, the review remains incomplete for the claim requiring them. Component tests and a runnable file do not prove that an instrument failed a complete research system, nor do website software tests create independent strategy-validation observations.