Predictive Analysis - Kicked Cars

Trac-K-icked

A predictive analysis on bad buys

Executive Summary

Over the past decade, used car market has been growing rapidly. When we walk to an auto dealer to make a used-car purchase, we take so many factors into consideration. So do the auto dealers while buying the same from an auto auction, a platform provided for auto dealerships to purchase used cars. Below are a few challenges faced by auto dealerships frequently with a used car purchase.

* Risk of unidentified mechanical problems and other issues which prevent from being sold to customers – such purchases being called ‘Kicks’ by auto community.

* Risk of incurring huge costs towards Kick repairs if bought

* Risk of unidentified true resale value of a used car because of which dealers make losses

The goal of this project is to develop a predictive model to predict if the car purchased is a Kick- bad buy based on some chosen attributes. This model can be used by dealerships while making their next purchase and minimize the risk of purchasing a kicked car.

While developing this model, we prepared data using the dataset of around 70,000 records with 34 attributes each which include information about the specifications of the car, about the auction and about price measure. The target variable is ‘IsBadBuy’ which indicates whether a car is a Kick or not.

We have observed certain data integrity issues such as incomplete data with several attributes of various rows being recorded as NULL and a few poor quality variables which have no effect on used car purchase. To minimize risk with such issues, data cleansing has been done using various techniques which brought down the dataset to around 40000 records on which various prediction methods are employed. While exploring different methods, we have considered 15 attributes as predictor variables with key factors being Vehicle age, Color, Make, Odometer reading, selling prices, Nationality etc.

The explored methods are evaluated based on...