Classification metrics versus trading value
Read accuracy, precision, recall and class imbalance without confusing a correct label with a profitable executable decision.
Topic notes and detailed research exercises.
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.
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.
Keep candidate selection, independent strategy validation, software checks and cloud production state as separate durable records.
A detailed experiment contract covering data availability, preprocessing, execution, costs, population and reserved evaluation; includes a blank JSON template.
Specify a causal target, inspect label boundaries, freeze entry identity and preserve experiments before measuring a trading implementation.
Identify the record scope, reconcile signals and open exposure, and separate cash flows, research fills and subscriber results.
Document decision inputs, timing and intervention authority so systematic and discretionary records can be compared without hindsight.
Track features, data, labels, selection, exits and release identity so a research result cannot silently become a different live claim.
Review a model as a versioned decision system: data identity, entry objective, exit ownership, reserved evaluation and durable operational state.
The expanded guide set covers model cards, Brier calculations, conditional gaps, thresholds and prevalence references. A second path follows label windows, as-of inputs, transformations, material revisions, drift and independent completion. Financial aggregation closes the chain by separating model summaries from account equity. The original build, track-record and audit guides remain available. Each added article owns a distinct evidence question and an invented case, not a near-identical vendor recommendation. The collection is a research-documentation library rather than a catalogue of unpublished proprietary model returns.
The workbook links three fully filled teaching cards, seven compact metric cases and the complete hundred-row default specimen. Existing planner exports and count JSON remain separate. Readers can compare actual object scope: document specification, probability arithmetic, threshold decisions and release evidence fields. No artifact supplies a fitted financial model, authentic market ledger or independently completed master. Downloading the files does not authorize a live strategy or convert hypothetical labels into broker executions. Follow the source maps for mechanisms and the limitations for what those sources did not verify.
If a group table hides varying predictions, begin with constant probabilities. If equality changes decisions, begin with threshold policy. If a source date is mistaken for arrival, begin with as-of availability. If a scaler fit touched evaluation rows, begin with frozen transformations. If a model name survived a material change, begin with validation scope. If the headline adds unrelated percentage results, begin with account weights. This question-driven path helps a reader find the specific failing boundary without forcing every problem into the same generic backtest checklist.