Classification metrics versus trading value
Read accuracy, precision, recall and class imbalance without confusing a correct label with a profitable executable decision.
No. Class prevalence, label construction, costs and the action policy matter. In the invented worksheet population, predicting the majority negative class always would already achieve 74% accuracy.
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.
No. Class prevalence, label construction, costs and the action policy matter. In the invented worksheet population, predicting the majority negative class always would already achieve 74% accuracy.
No. Its chronological splits and optional gap are useful mechanisms. Variable overlapping labels, revised data, external availability clocks and fit scopes need an explicit experiment-specific audit.
A registry records artifacts and lifecycle information. Strategy evidence and authenticated deployment/readback are separate records. Neither a file name nor a research LIVE label establishes production state.
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.
No. A deterministic rule can be systematic when inputs, timing, decision and exceptions are explicit. A fitted classifier adds parameter-estimation and transformation questions, not an automatic claim to superiority. Both approaches need a candidate registry and honest evaluation boundary. Calling a rule discretionary does not make it impossible to document; calling a model quantitative does not make its source causal. The filled cards distinguish these cases without inventing a head-to-head return study. A method becomes inspectable through its record, not through an algorithm family name alone.
No. Binary Brier loss averages squared differences between supplied label probabilities and actual zero-one labels. The lens requires a constant declared probability in each group. It does not infer probabilities from grades, fit a classifier or define what a financial win means. Its default labels and forecasts are invented. Zero gap in three groups is not independent evidence of future calibration. A future real study needs exact label ownership, sample coverage, frozen probability generation and evaluation provenance before it can make even a bounded label-forecast claim.
No. It excludes a declared number of decisions at each segment end under one fixed forward-label horizon. Variable targets, revised features, event arrivals and cross-asset joins need their own causality checks. Earlier observations can be permitted warmup while their later labels remain prohibited training material. A chronological sort alone cannot prove that every input was available. Inspect a complete example from original source version to transformation and label end. The planned count of eligible decisions is not the count of independent executed trades.
Provide a complete model identity, causal data contract, predeclared selection procedure, frozen entry and exit, full eligible record and independently scoped financial master. Follow with operational registry, deployed artifact, subscriber routing and certification evidence if a live product is claimed. Customer execution requires further account and timing records. A source can support one of these stages without supporting all of them. The release workbench is a place to collect questions and references, not a grading machine that creates financial certification from manually selected evidence states.