Skip to content

Validation Strategy for ASRQuant 1.3.0

ASRQuant separates software correctness, numerical correctness and research validity. Passing one layer does not imply the others.

Software gates

  • isolated domain test groups;
  • compile/import checks;
  • public API signature snapshot;
  • clean wheel/source-distribution installation;
  • hosted Python/OS matrix;
  • security and supply-chain workflow.

Numerical gates

The 1.3 suite includes deterministic invariants such as:

  • Black-Scholes put-call parity;
  • discount/zero-rate round trips;
  • exact par-swap repricing;
  • curve quote repricing;
  • quote/Jacobian finite-difference consistency;
  • risk/P&L reconciliation;
  • calibration recovery on known synthetic models;
  • covariance positive-semidefinite checks;
  • fixed-seed Monte Carlo reproducibility.

Research validation

asr.validation provides chronology-aware splits plus CPCV, PBO, Reality Check, SPA, leakage diagnostics and multiverse analysis. These are evidence-management tools, not proof that a strategy will remain profitable.

Data lineage

Use DataSnapshot, DataStore, PointInTimeFrame, Experiment and ResearchGraph when the exact information set and dependency chain need to be reconstructed later.

Scope discipline

A clean leakage report does not prove that a vendor dataset is point-in-time correct. A statistically significant multiple-testing diagnostic does not establish economic causality. A calibrated curve/model is not considered valid solely because the numerical solver converged.