Framework
One method, four labels. Every model uses a point-in-time frame, an observable target, time-safe features and walk-forward evaluation. This page owns that shared method. Pipeline owns handoffs; Data owns inputs; Contracts own obligations; Audits own observed results.
1. Prediction frame
At a fixed month-end T0, the model asks whether its target event will occur in the following six months. Features may use only information knowable at or before T0. The label may use only the future outcome window.
A positive is an event inside that window; a negative is meaningful only when the event could have been observed. Random K-fold validation is not valid for this frame because it mixes future and past records.
2. Labels and observability
The time frame is shared; each model owns its target definition. Permit models also need a coverage boundary so an absent permit is not mistaken for a true negative.
| Model | Positive event in T0+1..T0+6 | Owner of detail |
|---|---|---|
| roof | Roof-replacement permit | Model contracts and dated label evidence |
| garage | Any garage permit | Model contracts and permit classification rule |
| windows | Any windows permit | Model contracts and permit classification rule |
| rei | Property sale; dealable outcomes are a separate research target | REI research Log and evaluation contract |
Coverage: when a zero is trustworthy
Coverage is an observability condition, not a second label. A property can be treated as negative only for the source, jurisdiction and time range in which the relevant event would have been captured. The applicable contract states the obligation; Audits record what was observed. A draft or candidate is not an adopted rule.
3. Feature library
Feature families are shared, but each model selects and validates its own subset. Data owns the source fields and joins; the feature taxonomy owns the detailed inventory.
| Tier | Family | Boundary |
|---|---|---|
| A | Property physical | Parcel, use, age, area and lot characteristics |
| B | Owner and distress | Ownership, leverage and time-safe distress history |
| C | Valuation and activity | AVM, assessment, equity, appreciation and tenure |
| D | Date differences | No future records found in the completed audit; unusually strong dependencies still require model-specific ablation and monitoring. Evidence |
| E | National context | Release-aware macro series available by T0 |
| F | Local context | County or state context lagged to its publication date |
4. Validation and decision gates
Walk-forward folds train on earlier T0 cohorts and evaluate later cohorts. The embargo is at least the six-month label horizon, so train and evaluation outcome windows do not overlap. Rare-event sampling may reduce training cost, but calibration must restore the true base rate before a score is interpreted as a probability.
| Gate | Question | Evidence owner |
|---|---|---|
| Data | Did required contracts and fail-closed data checks pass? | Contracts + dated Audits |
| Label | Is the event definition observable for this cohort? | Model contract + coverage evidence |
| Model | Does held-out performance beat the predeclared incumbent? | Model audit |
| Release | Are calibration, caveats and output rules satisfied? | Adopted contract + release evidence |
Locked-test invariant. The March 2025 REI test remains untouched until Eduardo and Camilo sign off in writing. A proposal, candidate, draft, adopted contract and deployed implementation are distinct states; the owning source must name the state and date.