Acceptance conditions
An acceptance condition turns a research requirement into an explicit check: for example, whether a statistic is above, below, or within a specified boundary on a designated sample. The evidential value comes from knowing which rule was active, which data it used, and when the rule was chosen.
What can mislead you?
- Changing thresholds after reading results turns a confirmation check into another selection step.
- A passing condition does not establish robustness outside the metric and sample it evaluates.
- Metric names, cost treatment, and sample scope must stay attached to the rule.
Higher-precision retesting
A higher-precision retest holds the research question and strategy logic fixed while increasing the fidelity of execution assumptions, price-path handling, or other simulation details supported by the testing platform. The comparison is useful because a result that changes materially under a more exact model may have depended on approximation rather than robust behavior.
What can mislead you?
- “Higher precision” is not one universal setting; the exact model and data requirements are platform-specific.
- A stable result under one more detailed simulation still does not prove live execution will match the backtest.
- If strategy logic, costs, data, or sample boundaries change at the same time, the comparison no longer isolates execution-model sensitivity.
Calendar-year consistency
Year-by-year checks can reveal whether an aggregate result depends heavily on a small number of calendar periods. Useful summaries include the number of years represented and the share of those years with positive net outcome under one consistent accounting convention.
What can mislead you?
- Calendar years are arbitrary boundaries and are not themselves market regimes.
- Short samples or low trade counts per year can make percentages unstable.
- A profitable year can still contain severe drawdowns, concentration, or fragile subperiods.
Evidence and record chains
An evidence chain helps make a research result reproducible and auditable. Stable identifiers, revisions, hashes, or equivalent records can prevent later files or runs from being silently substituted for the artifacts that were actually evaluated.
What can mislead you?
- A strong record proves identity and lineage; it does not prove that the research method was economically valid.
- A complete chain cannot eliminate data leakage, selection bias, or weak assumptions elsewhere in the process.
- Current mutable system state should be distinguished from historical evidence captured at the time of a run.
Keep rule, sample, and evidence identity together.
A validation label is only shorthand. To understand it, retain the exact condition, the sample on which it was evaluated, the metric definition, and the evidence record that identifies the evaluated result.