Skip to model research
Systematic Trading ModelsSpecification / Evaluation / Release

Working glossary

An observation whose declared actual and predicted labels are both positive. The label definition and threshold must be fixed to interpret the count.

Original editorial revision: 9 October 2026. Expanded edition: 14 October 2026.

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.

True positive

An observation whose declared actual and predicted labels are both positive. The label definition and threshold must be fixed to interpret the count.

Balanced accuracy

Here, the average of positive-class recall and negative-class specificity. Undefined if either actual class is absent; not a measure of trading profit.

Label horizon

The future observation span used to resolve a target. Chronological boundaries must prevent training decisions whose labels reach into an evaluation region.

Artifact identity

A digest and version relationship connecting source, settings, data contract and outputs. An identity label does not independently validate the contents.

Forecast score, probability and action

A binary probability is a supplied value between zero and one for the declared positive label. A Brier score is the mean squared probability error on that label population; this edition uses the binary zero-one scale. An action threshold applies a separate comparison, with equality explicitly included in positive predictions here. Group fraction means observed positive labels divided by observations in a declared group. These terms concern labels and decisions, not cash payoff. A probability grade that was never calibrated cannot be relabelled a numerical win chance simply because a calculator accepts probabilities.

Calibration and conditional group gaps

Calibration describes the relationship between forecasts and label frequencies under a specified evaluation process. The lens's weighted absolute group gap is a narrower descriptive calculation for three constant-probability groups. It is not a fitted calibration map or an independent evaluation. A group with no observations has undefined observed fraction and undefined individual gap, and contributes no weight. Brier loss also reflects other aspects of forecast quality; a smaller loss alone is not proof that calibration improved. Real calibration needs an identified population and an honest fitted-versus-evaluated boundary.

Identity, master and deployment

Model identity joins code, settings, objective, source contract, target and transformation policy. A research master is the independently required completion artifact under the actual research method, not a synonym for an experiment log. Deployment identity identifies a cloud artifact actually served in a named environment. Subscriber routing identifies the path that a receiving account uses. Certification has its own scope and owner. Reusing a model nickname across materially changed versions does not make these records interchangeable. Missing artifacts remain missing even when the public site renders correctly.

Population, prevalence and raw aggregate

Population defines which observations entered a calculation and which were excluded or unresolved. Prevalence is the positive-label fraction in that population. The same-table reference assigns that fraction to every row; its Brier score is q times one minus q. It is not an independently frozen forecasting strategy. A raw aggregate return can mean a sum of separate percentage results and is not necessarily a portfolio equity return. Converting between the two requires dated exposures, account weights, cash flows, costs and compatible accounting. A glossary definition cannot reconcile an unexplained supplier total.