Skip to model research
Systematic Trading ModelsSpecification / Evaluation / Release

Research reading order

Read the confusion-matrix guide before interpreting classification scores. Then audit feature availability and transformations against chronological boundaries. Use the existing checklist and planner to write the experiment contract.

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.

Work from labels to decisions

Read the confusion-matrix guide before interpreting classification scores. Then audit feature availability and transformations against chronological boundaries. Use the existing checklist and planner to write the experiment contract.

Finish with artifact identity

The registry guide explains how to preserve candidate history, independent evaluation, software tests and cloud deployment as separate facts. A stored file is not a deployment certificate and a deployment is not a profitability certificate.

Library

A specification-first reading path

Begin with the three fictional model cards and compare what each actually declares. Follow the guides on as-of inputs and frozen transformations before judging any numerical score. Next use the chronological planner to make a label-boundary question explicit. Then inspect the probability packet and prediction-count worksheet. Finish with the release workbench, which asks for actual evidence references. This path moves from object identity to data ownership to metrics to release, instead of starting with an attractive score and inventing a method around it.

A probability-evidence reading path

Start with constant-probability groups and binary Brier score. The full default row specimen makes the compact table reproducible. Read calibration gaps next to learn why a group-average forecast cannot stand in for varying row probabilities. Compare the same-table prevalence reference and then alter the decision threshold. The equality case is deliberately visible. Finally read the financial aggregation guide: none of these label statistics includes execution or cash weights. A model that discriminates labels still needs an economic validation record under its native trading rules.

A deployment-evidence reading path

Inspect the model-version and validation-scope guide before relying on a live label. Review what a registry can store and what it cannot verify. Find the independently required master, then the actual cloud artifact and subscriber route. Drift monitoring belongs after the frozen deployment identity is known; it does not authorize silently retraining a selected model. These steps are documentation questions, not an instruction to initiate a new instrument study or purchase capacity. The site can describe a compliant operational chain while honestly marking missing proprietary evidence.

An honest stop condition

If data arrival, label ownership, fitting boundary or independent completion cannot be established, stop the stronger conclusion and state the precise missing artifact. Continue useful source review without presenting an intermediate zero as a full model failure. The downloaded workbook's fictional completeness does not solve any real supplier's missing record. Repetition across publications cannot create primary evidence either. The best reader deliverable is a clear question that its named artifact could answer, plus a bounded statement of what is currently supported. There is no need to fill an evidence gap with a newly invented percentage.

Model-card and probability references