data / features / feat.labels.pit
Labels · forward returns / triple-barrier feat.labels.pit
Pre-computed labels for ML pipelines. Triple-barrier, fixed-horizon, meta.
PIT · Pro tier · 2003-01 → today · 9.2B rows
Coming soon — not served by the API yet.
Methodology
A label is a statement about the future, so it is the one table that cannot be point-in-time in the ordinary sense — what it must be instead is exactly reproducible and clearly separated from the features it scores. Each row records the entry and exit timestamps actually assumed, the barriers in force, and the parameter set version, so a label cannot be read as if it were knowable on its decision date. Barrier touches on a path that gapped through the barrier are resolved conservatively, and the rule that resolved them is part of the version rather than an undocumented default.
Schema
| Column | Type | Role |
|---|---|---|
as_of_ts | timestamp | knowledge time · UTC |
effective_ts | date | PIT key · decision date |
symbol | string | ticker in force on effective_ts |
label_id | string | triple_barrier · fixed_horizon · meta |
horizon_days | int32 | maximum holding period |
barrier_up | float64 | upper barrier in returns |
barrier_down | float64 | lower barrier in returns |
entry_ts | timestamp | assumed fill |
exit_ts | timestamp | barrier touch or horizon |
label | int8 | -1 · 0 · +1 |
label_version | string | parameter set identity |
Use cases
- Train a classifier on triple-barrier outcomes.
- Hold labels fixed while features are rebuilt.
- Audit a training set for a horizon that overlaps the test split.
Background
- Look-ahead bias in quantitative trading: six mechanisms and their fixes
- Point-in-time market data: as-of dates, restatements, and adjustments
Related packages
- feat.tech.pit — Technical features · 142 cols