Public data index › Resolved permit forecast accuracy

Resolved cases only · reproducible evidence

What the permit forecasts actually got right

This track record joins the earliest sealed forecast for each permit to the newest qualifying assessor outcome. It reports the score we have, including the uncomfortable limits, rather than treating unobserved outcomes as wins.

Right-censored forecasts never improve this score. The current archive has matched assessor rows, but no model-case membership has reached even the conservative 12-month observation boundary. Every current rate is therefore UNKNOWN—not 0%, not a pass, and not evidence of positive or negative discrimination.

Horizon-resolved record

A hit means a probability at or above 0.50 matched a positive outcome, or a probability below 0.50 matched a negative outcome. Parentheses are the 95% Wilson interval. N is printed for every result; N=0 remains UNKNOWN.

19039 matched model-case memberships are still right-censored at 12 months and do not enter that denominator; 0 more lack provable pre-outcome ordering.

Permit-family detail

What was excluded

• intent score, 12 months: 18677 without a qualifying matched outcome, 14382 right-censored, 0 without provable pre-outcome forecast ordering, and 0 with an invalid probability.
• intent score, 24 months: 18677 without a qualifying matched outcome, 14382 right-censored, 0 without provable pre-outcome forecast ordering, and 0 with an invalid probability.
• intent score, 36 months: 18677 without a qualifying matched outcome, 14382 right-censored, 0 without provable pre-outcome forecast ordering, and 0 with an invalid probability.
• psir prob, 12 months: 7557 without a qualifying matched outcome, 4657 right-censored, 0 without provable pre-outcome forecast ordering, and 0 with an invalid probability.
• psir prob, 24 months: 7557 without a qualifying matched outcome, 4657 right-censored, 0 without provable pre-outcome forecast ordering, and 0 with an invalid probability.
• psir prob, 36 months: 7557 without a qualifying matched outcome, 4657 right-censored, 0 without provable pre-outcome forecast ordering, and 0 with an invalid probability.

A slice with zero horizon-resolved cases is printed as UNKNOWN and never carries a percentage. A positive also requires a forecast provably sealed before the year-granular transfer window. A matched parcel row is not automatically a resolved negative. The archive also contains multiple outcome vintages for some permits, so the raw outcome-row total is not used as N.

Method and provenance

Evidence effective date: 2026-08-01
Source: Los Angeles County Assessor parcel roll
Input-set row seal: 0df12e6c2b68a90120238281041ccd7827b7faca4f311dd0a540e686452de79a
Artifact seal: c83ef93e3be5d9d60cdc633708add2d35ab7936581fc37671f5b3c18dd8a47b8

1. Choose the earliest non-null forecast per permit and model family, with snapshot date and model version as the total ordering, and verify its stored row seal.
2. Choose the newest matched assessor vintage with both issue year and transfer-year proxy present, then verify its stored row seal.
3. Resolve a positive when its proxy year is strictly after the permit issue year, within the stated horizon, and the forecast snapshot predates January 1 of that proxy year. Same-year or later snapshots are chronology-unknown and excluded. Resolve a non-positive only when January 1 of the sealed roll year has reached the precise issue date plus that horizon, and only when the forecast snapshot itself is strictly before that horizon boundary.
4. Count hits and reproduce a two-sided Wilson interval from those resolved rows.

The assessor field is year-granular and can lag recent transfers. Same-year transfers are not counted as positive because they can predate the permit. This is a historical proxy result, not a promise of future accuracy or a sale-date record.