Classification metrics versus trading value
Read accuracy, precision, recall and class imbalance without confusing a correct label with a profitable executable decision.
Keep candidate selection, independent strategy validation, software checks and cloud production state as separate durable records.
Research design and arithmetic, not model recommendations. Classification scores are not monetary returns; a registered artifact and a passing software test are not independent strategy validation.
Give each candidate an immutable identity connecting source code, settings, data contract and parent experiment. Record the research objective and predeclared selection criteria. Preserve failed, aborted and revised candidates with reasons rather than removing them from the visible trial count. A registry created only for the winning run conceals selection. The goal is not paperwork for its own sake: a later reviewer should be able to answer which alternatives were considered, which sample was consulted and which exact candidate first saw the reserved evaluation. Those questions cannot be answered by a model name alone.
MLflow tracking documents runs, parameters, metrics and artifacts. Those facilities can help preserve experiment evidence, but an uploaded metric can still be incorrectly computed or based on a leaked population. Record the metric definition and denominator alongside the value. Attach independent evaluation receipts, causal-data checks and execution assumptions as their own artifacts. Do not infer validation from the presence of a registry stage or a successfully stored file. This resource note describes documentation; it has not benchmarked a private MLflow installation or verified a proprietary trading model with it.
Keep the entry objective and ranking basis distinct from the chosen exit and execution accounting. If the experiment predeclares one exit, preserve its version and do not open a new exit tournament after seeing the reserved result. A later entry or exit change is a new research identity, even when the marketing label is unchanged. Outcome statistics may validate or reject the frozen system without becoming a retrospective explanation for selecting the original entry. Save the native holding and risk policy so that a reviewer does not unknowingly compare two different systems under one familiar name.
A candidate marked LIVE in a research file, an endpoint listed in a document or code present on a branch does not establish that a subscriber route serves it. Separate research completion, code and merge state, registry admission, deployed source identity, routing and certification. An authenticated cloud deployment record should identify the actual artifact and environment. A readback should demonstrate the served version. Even that does not prove a future trading edge: it establishes a narrower operational fact. Version names and dates are useful references but are not substitutes for the deployed content or durable evidence.
For this workspace, production publishers, data refreshes, schedulers, monitors and verifiers belong in managed cloud services. Local research and temporary tests are allowed. A local cron job, tunnel or PC fallback is not a compliant production dependency. Durable cloud artifacts should preserve refresh status and the inputs actually used. A deployment plan that requires an operator's laptop to stay online has not completed that operational boundary. Keep cloud promotion separate from research admission and do not purchase extra capacity simply because a candidate deserves further testing.
Record why a version was paused, replaced or rejected and where the prior evidence remains available. A corrected metric needs its original result, replacement computation and impact on decisions. A rolled-back deployment needs the restored artifact identity and subscriber route evidence. Do not claim a model failed complete research because an intermediate component produced zero admitted candidates; the full declared method determines completion. The practical reader deliverable is a compact ledger linking each status to an artifact, not one vague green badge that mixes research success, software correctness, cloud availability and hypothetical financial performance.
Create four separate empty records titled research decision, software fixture, cloud deployment and public readback. List which identifier and evidence each would need. A passing count-formula fixture belongs in the software record, not the research record. A model file present on a branch belongs in code history, not the deployed-state record. Mark unavailable evidence as unavailable. The point is to prevent a single status word from claiming more than its receipt can support, before any real trading candidate is involved.
Documentation establishes the stated mechanism, not tested trading performance. Checked 9 October 2026; a later version can differ.
Read accuracy, precision, recall and class imbalance without confusing a correct label with a profitable executable decision.
Audit availability times, train-only transformations and forward targets before trusting a trading-model split.
The revised-contract card changes a feature availability rule and receives a new teaching identity. It deliberately carries no inherited validation, fitted parameters or deployed source hash. A real successor needs its actual source and allowed prior evidence scope, not a green status copied from its parent. Registry, research master, deployed cloud artifact, receiving route and certification remain different records. A monitor noticing drift cannot authorize retraining that changes those identities. Keep the full canonical completion process outside a simplified documentation fixture, and preserve every actual decision artifact under its original meaning.