Skip to system research
Build A Trading SystemRules / Evidence / Replay

Parameter search history and negative results

Retain the real choice set rather than presenting only the selected attractive curve.

A final candidate is not the search population

A research process can consider windows, universes, transformations, entries, exits, cost assumptions and reporting intervals. Those choices matter even when the final rule looks simple. Record what was actually evaluated and the reason for each rejection or revision. A short list of published winners is not the complete research population. Do not invent a candidate count or claim that an unrecorded search was small. Economic selection also differs from software correctness: a candidate that crashes may fail implementation checks without providing a market observation about the hypothesis.

The teaching fixture does not optimise

The sample uses one fixed 3/5 rule and one time exit over deliberately constructed observations. The library's optimiser is not called. Its negative 1.20 cash is preserved with all six transactions, not replaced by a setting chosen to create a positive screenshot. That establishes a complete single-fixture teaching record, not a financial search ledger. No unseen market evaluation, training period or calibrated objective exists. A reader changing a parameter can create another clearly labelled software exercise, but neither configuration should be marketed as an independently validated strategy discovered from real observations.

Classify the reason for a change

A source availability bug, arithmetic defect, failed economic hypothesis and operational route mismatch are different findings. A bug fix can invalidate an earlier result, while a properly tested negative market result can remain valuable evidence. Preserve the exact source/data/settings identity, exposed periods, observed defect and next action. A revised file hash identifies changed bytes but does not restore an untouched holdout. Keep old and new records available rather than merging them into a single favourable chart. A copied model nickname or publishing date cannot conceal the actual sequence of research decisions.

A bounded negative result stays useful

A complete negative finding under an authorised real research process can justify no model or no promotion, but it needs that process's required independent checks. A synthetic negative total is not evidence that a real instrument or family failed the entire factory. Conversely, a synthetic positive variation is not proof of a profitable cause. The publication adds neither a performance league nor a fabricated expert endorsement. Its purpose is to show the records that allow a later reviewer to distinguish selection, correctness, economic rejection and operational readiness without inventing missing periods, counts or audit scope.

Filled invented system case

The fictional decision log has one declared configuration, no optimiser call and six completed simulated transactions. Total net cash is negative 1.20. A proposed change to the rolling windows is recorded as another fixture revision, with the earlier result retained. There is no actual candidate family count, independent financial holdout or verified no-model master. The completed teaching conclusion is that the original software exercise is reproducible, including its unfavourable total.

ObjectRecordBounded finding
Choice setOne fixed fixture ruleNo optimiser
Observed fixtureSix simulated positions-1.20 net cash
Proposed revisionNew source/settings identityOld record retained
Financial rejectionFull actual process requiredNot inferred
Invented closes and moving averages above simulated equity
Invented library replay, not historical data or economic validation.

Primary references and scope

Expanded source review: 13 October 2026. Documentation supports the stated mechanism, not a tested account, hypothetical fill or provider outcome.

Related research questions