Upward-Move Scorecard
Credit-scorecard methodology applied to price direction: daily trend features are coarse-classed into bins, converted to Weight of Evidence, ranked by Information Value, fitted with logistic regression and scaled into points — the same pipeline used to build a credit scorecard.
Read the numbers with credit-scoring instincts inverted.
A credit scorecard separates a rare event (~3% default) from a common one, so IVs of
0.1+ and AUCs of 0.75+ are normal. Predicting whether a stock rises is a near coin flip
(~50% base rate), so IVs here are one to two orders of magnitude smaller and an
out-of-time AUC of 0.53 is a genuine result, not a broken model. The split is strictly
by date, with a gap so no training row's outcome overlaps the test
window — a random split would leak the future and produce a flattering, false AUC.
No scorecard fitted yet. Run
docker compose exec web python manage.py fit_scorecard.