Buyers of used cars have to predict their dealing prices at auto auction held afterabout one month, but it is very difficult to predict them because each condition of used cars iscompletely different such as mileage, model year, body color, etc. For this reason, we proposetwo prediction methods: as the first one, we consider the median of dealing prices in each carmodel as a base price and predict its future price by time-series models: ARIMA and SARIMA.After that, the predicted bace price is converted into the individual price of each used car by themachine learning method that learned the relationship between the condition of used cars includingindividual prices and the bace price. As the second method, we adopt the deep learning approachto directly predict individual future prices of used cars without using the base price, but using allthe information attached to each used car as explanatory variables. To verify the usefulness of ourproposed methods for a used car assessment system, we performed some prediction tests using thereal auto-auction price data.
