Demo 2: factor features and separate evaluation

Before you begin

This tutorial records the unpublished main source candidate 0.2.0rc1. Bring your own local data and freeze it using the shared preparation guide. Results remain RESEARCH-ONLY with PIT guarantee false.

Shared preparation: select and freeze data

Freeze the factor recipe

In “Data Center → Factor diagnostics · RESEARCH-ONLY”, choose the same frozen manifest. Fix 20- and 60-session momentum, positive direction with 0.5 weights, rank normalization, minimum cross-section 3, three quantiles, and a five-observation forward label before running.

Read features, labels and diagnostics

Check the feature/recipe page first. feature_rows are computable through the signal date’s D-1 close and are stored separately from evaluation labels. A label is the simple return from execution open to the fifth subsequent observed-session open.

It belongs only to the evaluation panel; unavailable tail labels remain missing and are not strategy features. Review daily IC/Rank IC, valid cross-section and coverage, quantile returns, quantile membership turnover, factor correlation, and descriptive year summaries.

Constant factors, missing sections, and invalid dates must show a reason. Code-order tie breaking is deterministic display, not information. Three securities yield at most three quantiles; overlapping labels do not create independent observations.

ICIR is a descriptive mean/std measure, not calibrated inference.

Keep evaluation separate from strategy inputs

Never feed forward labels, quantile returns, or the evaluation panel to a strategy. Do not choose direction or weights from this historical window. Keep contract.json, receipt.json, report.json, and run logs for reproduction. If dates or coverage are incomplete, diagnose the receipt instead of dropping a security to make the panel look complete.

Continue learning

All five tutorials · Quantitative research basics · Reading result charts