Skip to model research
Systematic Trading ModelsSpecification / Evaluation / Release

scikit-learn and MLflow: a model-research source map

Official metric, preprocessing, splitting and tracking documents with a bounded reader exercise.

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.

Read metrics as functions of declared counts

Use the confusion-matrix documentation to confirm actual-label rows and predicted-label columns. Reproduce the invented 18/12/8/62 worksheet population, then remove one actual class and inspect undefined denominators. This exercise checks arithmetic rather than trading performance.

Review the transformation boundary

The common-pitfalls guidance and TimeSeriesSplit reference concern fit scope and chronological mechanics. Build a tiny fixture where a transformation must fit only training rows and a five-observation label cannot cross the region end. A passing fixture does not prove that a real source arrived on time.

Store the actual experiment contract

MLflow can record parameters, metrics and artifacts. Attach the blank backtest contract after filling it with documented assumptions, then retain revisions and rejected candidates. A logged result is not independently verified merely because the storage software accepted it.

What is not assessed

No private installation, paid licence, operating benchmark or live strategy outcome was tested for this note. Pin the version actually used in a future experiment. Hosted access, account settings and operational security require their own current documentation and evidence.

Source documents

Documentation establishes the stated mechanism, not tested trading performance. Checked 9 October 2026; a later version can differ.

Probability sources added to the model source map

The Brier-loss and calibration references are appropriate for binary forecast definitions and their limitations. Our lens uses the zero-one binary convention with constant probabilities per declared group. It does not call a calibration training procedure or use a public library example as genuine market evidence. The expanded row file is authored solely to make its exact aggregate formula inspectable. A future research environment should pin its actual dependency version and positive-label convention. Different binary or multiclass scaling choices must not be mixed in a score comparison.

Storage tooling cannot settle validity

Tracking systems help retain parameters, metrics and artifacts, but storage acceptance does not prove population completeness, causal joins or independent master completion. Keep the source, selection and validation evidence linked rather than treating an experiment dashboard as an audit. The cards distinguish a written educational specification from an actual fitted run and a deployed route. A resource description can identify useful software without claiming a paid benchmark or a verified proprietary experiment. Preserve actual license, environment and cloud-operation requirements before using those tools in real production.

Read mechanisms beside their missing evidence

The planner exercises chronological row arithmetic; the count worksheet exercises denominator conventions; the group lens exercises probability loss. Their correctness fixtures do not multiply independent strategy samples. Actual instruments need the canonical research route and a full identified native exit and fill convention. Sources reviewed for this edition are documentation, not complete customer account records. Avoid importing legal, financial or operational guarantees from a source that addresses only a calculation function. When a document version changes, record the relevant difference rather than silently implying every previous experiment used the new behavior.