Data & features
From raw activity to comparable signals
The same rules explain the recurring suffixes in the feature names and keep signals comparable across very different projects.
Trailing time windows
A numeric suffix is a window ending at the observation time t: _30d, _90d and _365d count events in the last 30, 90 and 365 days, while _at_obs is a state measured exactly at t.
Only the past is used
Every value is computed from information available at or before t. Nothing after t enters a feature — that keeps the setup honest as an early-warning prediction rather than hindsight.
Relative, not just absolute
Where a raw count would not compare fairly across project sizes, the feature set adds a ratio, a share, or a year-over-year change — a large and a small project can be equally well or badly maintained at very different volumes.
Steadiness over bursts
Signals like active commit months and the activity drop capture whether work is sustained and whether it is decelerating, not only how much happened.