Five commitments for pipelines that feed models from staggered-arrival data.
If the entities in your data report on different dates — companies filing statements, clinics reporting cases, agencies publishing revisions — then the panel you train on is assembled out of records that arrived over time. Almost every leakage bug in that setting comes from forgetting it. The pledge below is a self-declared statement that your pipeline does not forget it.
This is a pledge, not an audit. There is no gatekeeper, no application, no review. You read the five commitments, decide whether your pipeline meets them, and add the badge if it does — the same way projects adopt Contributor Covenant or a code-style convention.
1. As-of joins everywhere. No value is read before its availability timestamp. Why: a join on the period a value describes, rather than the moment it became knowable, silently imports the future into every row it touches.
2. Arrival timestamps are kept, not discarded. Every record carries when it became available, not only what period it describes. Why: you cannot audit what you did not record — once the arrival time is dropped, no later analysis can reconstruct what was knowable when.
3. No silent restatement backfill. Revised values enter as new vintages; they never overwrite the value that was originally published. Why: overwriting history makes yesterday's backtest unreproducible and quietly upgrades the past with knowledge nobody had at the time.
4. Completeness is gated, not assumed. Cross-sectional and derived values state how much of their group had arrived when they were computed. Why: a cross-sectional quantity computed before the group finished reporting is estimated on a partial population, and a correct as-of join over a partial population is still a correct join over a biased sample.
5. Leakage is screened, not presumed absent.
Derived signals get a susceptibility screen before release — any tool; ours is
pip install pit-release-gate.
Why: "we used point-in-time data" answers whether values were readable, not whether the
set of entities that had reported was a selected sample. Those are different failures with
different fixes.
- Read the five commitments and check them honestly against your pipeline.
- Copy a badge snippet from badge-snippets.md into your README.
- Optionally link the badge to a short note in your own repo saying how each commitment is met — that is what makes the badge worth reading rather than worth scrolling past.
There is nothing to submit and no one to notify. If you later find your pipeline does not meet a commitment, remove the badge. Questions are answered in FAQ.md.
The completeness and arrival-selection commitments (4 and 5) are developed in three publicly available preprints:
- Wu, K.-T., & Wu, K.-I. (2026). Correct-by-Construction Factor Computation: A Verifiably Point-in-Time Engine for Tradeable Signals. figshare. doi:10.6084/m9.figshare.32952482
- Wu, K.-T., & Wu, K.-I. (2026). Measuring Incomplete-Cross-Section Leakage: A Matched Placebo, a Susceptibility Screen, and Evidence from Taiwan and US As-Filed Data that the Channel Is Benign and Correctable. figshare. doi:10.6084/m9.figshare.33061955
- Wu, K.-T., & Wu, K.-I. (2026). Susceptibility-Graded Release Control: Preventing Incomplete-Cross-Section Leakage in Financial Machine-Learning Pipelines without a Blanket Timeliness Penalty. figshare. doi:10.6084/m9.figshare.33158615
The broader leakage literature these build on:
- Kaufman, S., Rosset, S., Perlich, C., & Stitelman, O. (2012). Leakage in data mining: Formulation, detection, and avoidance. ACM Transactions on Knowledge Discovery from Data, 6(4), 1–21. doi:10.1145/2382577.2382579
- Kapoor, S., & Narayanan, A. (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9), 100804. doi:10.1016/j.patter.2023.100804
Maintained by Max Well Apex LLC. Text of this pledge is CC BY 4.0 — copy it, fork it, translate it, no permission needed.