3,2,1 – Gone! The Science of Used Car Pricing
Despite the recent Volkswagen scandal, German cars still have a world-class reputation. And Germans still love cars: more than 44 million vehicles are registered, that’s about one car for every second citizen. Understandably, the used car market is also large – but, how does one arrive at a fair value for a used car?
For a long time, arriving at a fair valuation for a used car involved taking a physical trip to a local car dealer or trader. After thorough inspection, one would receive an offer at which the dealer would be optimistic, to receive an adequate margin in the market. If not satisfied with the initial offer, one could re-negotiate or visit additional dealers to compare for a fair value (or try to talk friends and family into a deal). Of course, online-pure approaches have already reshaped lots of the old dynamics of the industry.
Especially a Berlin-based startup – the Auto1 Group – which operated “under the hood” for a long time, is now offering its customers to give a price estimate for the used car online, without having ever inspected the car. With the offer, the seller can approach the local representative to seal the deal. As online becomes more and more relevant as a channel, giving reliable price estimates based on a few data points becomes essential for every market player.
What are the ‘drivers’ of used car prices?
Without any knowledge about what the value components in cars are and how the car’s value decreases after the initial purchase, the first approach is to evaluate the questionnaire, which the Auto1 Group uses to arrive at a fair price, to get a rough understanding.
For each car model, the company requires the following data input:
- Date of registration: When has the car first been registered – this is essentially the age of the car
- Category: Within each car model, there are usually various categories, e.g. cabriolets, estate cars or limousines.
- Type: This gets a little more specific, as it refers to specific equipment of the car, as well as gear and style components.
- Mileage: simply the amount of kilometers, the car has been driven at the given time
These are surprisingly few data points, which – obviously – seem to give Auto1 sufficient confidence to arrive at a somehow realistic price point. Without diving too deep into the details, one thing seems clear: as cars have an incredibly broad variety of individual equipment, key common drivers are primarily age and mileage – two very basic aspects of the state of the vehicle.
Sounds easy? It’s probably a little more complicated
Using basic information about vehicles probably allows for a ballpark-like price estimation, but this is a rather isolated view on pricing dynamics, which are most probably embedded in a more complex world.
What happens to the market price for a car model, if it was recently announced that a new and updated version of that model will go on sale in the near future. The price will most likely go down (if the price of the new model is not outlandishly higher than the previous version).
However, a prediction model will be rather slow to adopt these market dynamics. Trained on vast amount of previous transaction data, the model has learned multidimensional correlations between features and the realizable market price point. If all of the sudden, the market price drops due to the release of the newer version of the model, the prediction model will continue to suggest high prices for these models.
Online Marketplace, but geography remains important
Another characteristic of the used car market is that it remains highly tied to geographics. Whoever sells the car will most likely only realize a price point, which is comparable to similar car models not further than 200km away. If one plans to sell a car in Munich, the prices for that car in Munich are more important than the prices that could be realized for the same car model in Hamburg. For example, a certain car might be high in demand in Hamburg, while in Munich there is rather low demand and currently high supply of the certain model by local dealers. These dynamics can shift the market prices, but prediction models might fail to acknowledge them sufficiently.
With more and more data availability, platforms like mobile.de, auto1 and autoscout24 have the opportunity to build machine learning algorithms, which bring more light into the hidden dynamics of the used car pricing. Focussing on the hard facts of the vehicle will be the first stage, but more fine-grained models will certainly have to account for within-category substitution and geographic effects. Determining the fair value of a used car with few data inputs though, is certainly getting within reach and slowly bringing more transparency and choice to the market.