Model card
How the price estimate works
Every listing gets a market estimate and an 80% range. The estimate comes from a gradient-boosted model (LightGBM) trained on average US asking prices per make, model and model year, with mileage and miles-per-year as features. The range is a split-conformal interval calibrated on held-out data, so roughly 8 in 10 true prices land inside it.
Model 20261118-1149-936dfae482dc, trained 2026-11-18T11:49:54+00:00 on 7,489 make/model/year groups.
Retraining
- A weekly Kubernetes CronJob downloads the latest data, retrains, and evaluates on a held-out split.
- The new model is promoted only if its error is no worse than 2% above the current one and its interval coverage stays within 5 points of 80%.
- The API reloads promoted models without a restart and exports the loaded version and test error to Prometheus.
Photos
Uploaded photos go through CLIP ViT-B/32, exported to ONNX and quantized to int8 (84 MB, about 28 ms per image on two CPU cores). Zero-shot prompts check that the photo shows a vehicle and suggest body style and color. On a small labeled set, body style was right first time on 75% of photos and in the top three on all of them; brand recognition was too unreliable to ship, so make and model come from the VIN instead.
Limits
- Training data are average asking prices, not sale prices, and don't see condition, trim or history.
- Rare models with few listings get wider, less reliable ranges.
- Auctions on this site are simulated and bids are not binding.
Data: ForCar open data, CC BY 4.0.