Frequently Asked Questions

Our primary ARV is produced by a market-trained 11-expert ML ensemble (XGBoost, CatBoost, LightGBM, ExtraTrees, DRF + quantile & uncertainty models) wrapped in a conformal prediction layer. The ensemble runs per market and is retrained on recent sold comps; the conformal layer emits an empirically-calibrated confidence interval (CI) based on held-out error.

The CI you see on an analysis is not a Monte-Carlo noise band — it’s the observed spread of the model’s residuals on validation properties it never saw during training, so the stated coverage reflects actual market-specific behavior.

That amber banner means a trained ML ensemble was not available for your property in that market at analysis time, and Recon fell back to a market-level heuristic (typical price-per-square-foot in the subject ZIP, scaled by the property’s square footage and adjusted for condition and year built).

Common triggers:

• Market not in our trained fleet. We train per-market; brand-new market coverage ships with a heuristic placeholder until enough local sold data accumulates.
• Model artifacts stale or loading. If a market’s retrain failed or a rollout is mid-flight, the ensemble can temporarily fall back to heuristic.
• Insufficient comparable sales. If fewer than the minimum threshold of recent nearby comps exist, the conformal calibration skips and we surface a heuristic rather than emit an untrusted CI.

When you see the banner, the ARV number, floor, ceiling, and CI are all approximations derived from market-level statistics — not conformal intervals. Please treat them as directional and verify with your own comps before underwriting.

The CI is an asymmetric conformal prediction interval calibrated against the specific market’s out-of-fold residual distribution (using a Mondrian local CQR estimator). An 85–90% CI means that, across the training market’s held-out data, roughly 85–90% of actual sale prices landed within the bounds we show — the exact target varies per market based on local calibration.

Some markets have smaller holdouts and get soft-blended with peer markets to stabilize tail coverage; you’ll see a “calibration” badge next to the ARV whenever that blending is active.

As of April 2026: Huntsville AL, Apex/Cary NC, Augusta GA, Birmingham AL, Greenville/Spartanburg SC, Charlotte NC, Chattanooga TN, Jacksonville FL, Savannah GA, Knoxville TN, Greensboro NC, and Cincinnati OH. Addresses outside these metros analyze with a heuristic fallback.

Data-quality note: a few markets (currently Huntsville and Augusta) have partial sale-date coverage upstream, which can widen their prediction intervals versus other markets. We surface deals in those markets conservatively while we backfill the missing data.

Every package reservation runs inside a Firestore transaction. If the downstream SNIPER pipeline times out or errors, a Cloud Function sweep detects the failed deal_packages doc and automatically refunds the credit on your monthly usage counter — you’re only billed for deliverable work.