How we forecast collector car values

The single source of truth for prices is OldCarsData, which aggregates Bring a Trailer and Cars & Bids auction outcomes (sold price, reserve status, mileage, condition). Catalog metadata — year, make, model, trim, engine specs — comes from CarQuery and NHTSA. We do not use dealer asking prices as primary inputs.

All forecasts are anchored to Condition #3 (Good) — a presentable, drivable example with no major flaws. From that base we apply segment-aware multipliers for #1 Concours, #2 Excellent, and #4 Fair grades. Multipliers are stored in condition-multipliers.json — never hard-coded.

We train an XGBoost regression stack with 1-year, 3-year, and 5-year horizons. The 5-year is a meta-model that consumes out-of-fold predictions from the 1y and 3y models as additional features. We validate using time-series cross-validation (scikit-learn TimeSeriesSplit) so we never leak future data into past predictions.

Inputs include trailing 12-month auction medians, reserve-met rate, segment trends, community demand (Reddit + Google Trends), production rarity, and macro correlates (S&P 500, gold).

From the model we derive three projections:

• Pessimist — the lower bound of the prediction band.
• Moderate — the central forecast.
• Optimist — the upper bound.

High / Medium / Low based on auction history density (more comps ⇒ more confidence), peer-group strength, and signal agreement between the three horizons. Vehicles with fewer than 5 auction results in the trailing 36 months are flagged insufficient data and we withhold a projection.

The recommendation compares forecasted CAGR against a hold-cost threshold (storage, insurance, opportunity cost). Above threshold ⇒ Buy. Below ⇒ Sell. Within band ⇒ Hold. Ultra-rare vehicles (< 100 confirmed production units) are never assigned a Buy signal — too illiquid to forecast reliably.

North American market only. Salvage, flood, and rebuilt titles are filtered from the default view via a denylist applied at load time.