Can an algorithm really learn to price used cars, a task that has traditionally relied on an experienced salesperson's intuition?
We spoke with Senior Advisor Mika Laukkanen about a project where skepticism turned into success. He shares how a machine learning model became the sales team's most important tool.
Pricing used cars
Who are you and what do you do?
I'm Mika Laukkanen, and I work as a Senior Advisor at Data Design. In practice, I take part in various data and AI projects.
You have extensive experience using machine learning and AI. Is there a project that has stuck with you?
That's a tough question, there have been quite a few projects over the years. Maybe pricing cars with machine learning is one of the most memorable ones.
Why?
It was a time when solutions like this were just emerging, and very few of the projects that got started actually made it into production. This was one of the ones that did.
What was the project about?
The goal was to use machine learning to price used cars. The idea was that it would save salespeople time, standardize pricing, reduce pricing errors, and improve sales efficiency. As an added option, the predicted prices made it possible to estimate inventory value and how it develops over time.
What kind of solution did you end up with?
So the project built a predictive model based on historical prices and the information available about the cars. The end result was an application where the salesperson enters a license plate number, and the app returns an estimated selling price and, for example, the predicted sales time.
What was the reception like for the solution?
You could say the reception was mixed among salespeople at first. For example, they quickly noticed when the model returned clearly skewed predictions, which was useful feedback for improving the model.
This also made it very clear that there is no such thing as an objectively correct price, meaning the same car can be priced in different ways. And ultimately this can also depend on the salesperson's own experience and perspective.
In any case, using the application has become an everyday tool in their work.
What kinds of challenges arose in the project?
What stuck with me most was that car prices change quickly. We had training data spanning a long period, but the older prices were no longer comparable to the present day. We had to find and test solutions for this.
Any other interesting observations?
There was clear variance in prediction accuracy between car brands. Let's put it this way: for so-called quality brands, the predictions worked with a smaller error. In addition, newer cars get more accurate predictions than older ones, which is a fairly predictable characteristic.
Could the same solution concept be applied elsewhere?
Yes, definitely. I believe a similar concept could be applied to pricing homes and many other products. The main thing is that we have enough high-quality data.
What was the best part?
The fact that the results went into production and further development. That's when you know something has gone right.
