Selling price suggestion with linear regression

There comes a time when you start noticing all these useless items around you – whether it’s because you badly need money to buy something new and shiny, or simply because you realized your apartment’s size has mysteriously decreased by 200 sq ft in the last few years. In moments like these, I usually want to sell a lot and do it as quickly as possible, meaning I would rather go for a fixed price instead of launching a 14-day auction. And here’s the tricky part: how to choose a reasonable price for my treasures? One way is to dig through similar listings, set up various parameters filters and then finally calculate the average market value for each of the items, but there are two problems with this approach:

a) Only ongoing listings are available, which means the real value of an item is often obfuscated – majority of the auctions were most likely already finished and their selling prices are unknown. There’s a risk that the listings we can see are significantly overpriced and that might be the main reason they’re still out there.

b) It’s slow, boring and tiring.

What if there was a much faster solution? I wish Ebay or its Polish competitor Allegro had a feature to speed up this process by supporting amateur sellers like me with a suggested price.

The idea is to create a prediction model for this particular task and automate it to some extent. Having access to huge amount of historical auctions data, one could create supervised learning algorithms for several categories. I tend to think it’s a feasible plan for items whose parameters can be unambiguously determined and represented by numerical values, e.g. cars, cameras and many other electronic devices.

The price is a continuous output, so by definition it’d be a linear regression problem with multivariable hypothesis function. Since I’m currently trying to get rid of a notebook, I’ll use it as an example. The model representation could be based on seven features that I care about (in reality, I’m definitely more picky than that): screen size (in), RAM (GB), CPU clock rate (GHz), number of CPU cores, disk size (GB), weight (kg) and release year. This gives us eight variables as a starting point, including the conventional intercept term \theta_0 x_0.

After researching the topic for a while, I found out that Ebay has indeed bought a company doing just that (and even more exciting things, like predicting the results of ongoing auctions) two years ago, so hopefully this feature will get real soon.