Methodology

The weekly universe

Each week, the pipeline rebuilds its candidate universe from scratch: every stock with a real technical trigger that week and a clean data-quality verdict. No fixed reference library, no matching against a fixed set of past examples.

The model

A machine learning model trains fresh every single week against a rolling window of recent history, then is discarded the moment that week's ranking is done — there is no persisted version carried from one week to the next. Four independently-trained models, each looking back a different length of time, vote together on every candidate; their averaged score is what ranks the week's candidates.

A candidate isn't published just for scoring well against its own week's field — it also has to clear a rolling confidence bar, measured against the trailing year of scores. Most weeks all 3 slots fill; some weeks fewer candidates clear the bar, and fewer get published.

Outcomes and the 12-week window

A candidate's outcome is not written until 12 weeks have elapsed. The metric is maximum favourable excursion (MFE) - the highest return the stock reached at any point within 30 weeks of the scan date. Outcome files are written once and never edited, enforced at the code level.