Osman-Ozcanli/car_price_prediction

Car Price Prediction (US Used-Car Auction)

LightGBM regression model that estimates the sale price of a used vehicle from a small set of attributes. Continuously retrained from validated user feedback as part of the `car-price-mlops` pipeline.

Intended use

• —Rough estimation of fair market value for a US used vehicle.
• —Demonstration of an end-to-end MLOps loop (feedback → validation → retraining → conditional deploy) — the model is the artifact, the system around it is the point.

Not intended for: legally binding valuations, insurance settlements, or markets outside the United States.

Inference pipeline (frozen)

input_dict → AddInteractions (age * odometer)
           → preprocessor (StandardScaler + OrdinalEncoder + TargetEncoder)
           → LightGBM model
           → PowerTransformer.inverse_transform (Yeo-Johnson)
           → clip(500, None)
           → × 1.38 inflation multiplier (2015 → 2025)
           → estimated price

Features

age_x_odo = age × odometer is added at runtime by the AddInteractions transformer. seller is intentionally absent — it was dropped after evidence that the production app could not reliably populate it.

Artifacts in this repo

AddInteractions is stateless and is constructed at inference time — it is not pickled or downloaded.

Training data

• —552,941 US used-car auction records (2015 vintage, public Kaggle dataset).
• —90/10 train/validation split; the same split is used to compare a candidate model against the active one before deploy.

Performance

• —Test RMSE: ~$1,814 on the 2015 validation slice (raw, pre-inflation).
• —Inflation adjustment of ×1.38 is applied at display time only — never during training.

Best hyperparameters

{
  "learning_rate": 0.01279,
  "num_leaves": 265,
  "min_child_samples": 15,
  "subsample": 0.671,
  "colsample_bytree": 0.700,
  "reg_alpha": 0.000455,
  "reg_lambda": 0.000275,
  "n_estimators": 3298
}

Tuned with Optuna over 50 trials.

Limitations

• —Trained on 2015-era data; market structure (EV mix, post-pandemic supply shocks) is partially captured by the inflation multiplier but not learned.
• —Sparse states (e.g. AK, ND, WY) fall back to the global mean via TargetEncoder — predictions there are coarse.
• —Body type Pickup is not represented in training and is excluded from the UI selector.
• —The model is recalibrated only when feedback accumulates — there is a lag between market shifts and updated parameters.

Versioning and rollback

Each successful deploy is tagged vYYYYMMDD on this repo. The previous active model is snapshotted to lgbm_tuned_prev.pkl before each deploy, enabling A/B labelling at inference time and a rollback target.

License

MIT. The training data is from a public Kaggle dataset; check the original source for any attribution requirements before commercial use.