How these reports are counted

Counting rules

• Each prediction is evaluated once at a declared horizon.
• Risk modes are reported separately. The headline is sample-weighted, not the best mode.
• Training or validation accuracy is shown only when a labelled holdout set exists. Otherwise it is “Not available”.
• Trades are paired BUY then SELL on the same symbol. Fees are subtracted from ROI.
• Reports that fail sample or quality gates stay in Historical with a label. Poor performance is not a deletion rule.

Publication gate

A current Insights, AI Assistant, or Trading report needs at least 30 scored observations, valid timestamps, and required fields. Historical pages can be smaller; they are labelled.

Limitations

• Overlapping windows are labelled; they are not independent samples.
• Dataset B (quality score 0) is excluded from trading headlines.
• Fees, pairing rules, and the 15-minute evaluation horizon are disclosed on each generated page.