Car prices prediction

Project description

Our customer is a used car sales service that developes an application in order to attract new clients. This app helps to determine a market price of user's car.

Following criteria are essential for the customer:

• prediction quality
• time required for model trainig
• time required for prediction

Dataset description

Dataset objects are entries crawled from car profile forms:

• DateCrawled — date the form was downloaded on from a database
• VehicleType — car body type
• RegistrationYear — year of car registration
• Gearbox — gearbox type
• Power — horsepower (hp)
• Model — car model
• Kilometer — car mileage (km)
• RegistrationMonth — month of car registration
• FuelType — fuel type
• Brand — car brand
• Repaired — was the car in repair or not
• DateCreated — car profile creation date
• NumberOfPictures — number of car photos
• PostalCode — postal code of a user who owns a car profile
• LastSeen — last user activity date

Target feature:

• Price — price (euro)

Conclusion

• Dataset has a lot of outliers and missing values.
• • I marked outliers as NaN values and removed only objects with 3 or more missing values.
• • For objects left I changed NaNs to unknown values in categorical features.
• All models except dummies showed RMSE score that meets reqirements.
• Random Forest Regressor showed itself to be the slowest one.
• Linear Regression was the fastest to fit and to predict. It has the worst RMSE score but it still fits requirements.
• CatBoost fits much slower than LGBM but makes predictions faster.
• LGBM fit time is the second after Linear Regression

LightGBM model was chosen with RMSE score of 1429