Accuracy (v0 — published baseline)
The numbers are the brand. This page reads /api/metrics/scorecard live; below is the current baseline.
Comp-engine backtest (leave-one-out AVM, 744 real arms-length 2-4 unit sales, Logan Square/Avondale 2022–24, no future leakage):
| Metric | Value |
|---|---|
| Median absolute % error (MdAPE) | 20.1% |
| Within ±10% / ±20% | 28% / 50% |
| Bias | +4.7% (over-values) — haircut model ARVs accordingly |
| Confidence calibration | Weak: conf 80–100 is only ~2–3% better than 40–59 |
What this means in practice: comp ARV is a ±20% screen, not an appraisal. The dominant error source is property condition, which the current engine cannot see — condition-stratified comps are the #1 improvement (P1). Guards added 2026-07 (similarity floor, dispersion gate, uncapped dispersion penalty, tract-anchored income cap) prevent the worst failure mode: a mixed-condition median mislabeled as ARV. When the gates refuse, the tool says “insufficient comps” instead of inventing a number.
Deals graded vs model: live from /api/metrics/scorecard (your tracked deals, predicted vs realized). As-of 2026-06 backtest; refreshed when the P1 harness lands. Model version: p0a-2026.07.