github.com/Zhanyl-tech/research-platform · Python · MIT
partial Phase 1 of 6 is built: the point-in-time data layer. Feature lineage (phase 2), the data-quality and leakage monitor (phase 3), experiment tracking, a walk-forward evaluation harness and scale-out are planned, not built. The reference data is a deterministic synthetic fixture.
You test a signal. It works. You put it into production and it does not.
Often nothing is wrong with the model. The database is a current-state
database, and it quietly told you things you could not have known: the revenue
figure was restated in November, your universe contains only companies that
survived, your prices are adjusted for splits that had not happened yet, and
the ticker ZZZ is two different companies spliced at the seam.
None of these throw an error. They produce a clean, plausible series and a Sharpe ratio that does not survive contact with reality.
Every record here carries the date it became knowable separately from the date it describes, nothing is overwritten (the API has no update or delete path), and every query takes an as-of date. That makes this class of bug structurally impossible rather than carefully avoided, with one precondition the store cannot check for you: every source must stamp its knowledge dates honestly.
make demo runs offline, with no credentials, and walks five traps
(survivorship, ticker recycling, a restatement, a corporate action, late data),
each showing the same query returning different and correct answers on
different as-of dates.
Phase 3’s leakage detector is the reason for the order: it only means something once the data layer can say what was knowable when.