New AI Trick Helps Online Marketplaces Set Better Prices

Imagine you run a website that connects buyers and sellers, like eBay or Uber. You want to set a price to buy from sellers and a price to sell to buyers, earning the difference. But you don't know the lowest price a seller will accept or the highest price a buyer will pay. You only see if a trade happens. How do you learn to set prices that maximize your profit? That's the problem researchers tackled.

They developed an algorithm that, over many trades, adjusts prices based on whether trades succeed or fail. Surprisingly, they found that learning both sides of the market is no harder than learning just one side. The algorithm's performance, measured by 'regret' (how much profit it loses compared to the best fixed prices), grows at a rate of T^(2/3), which is the same as the classic result for one-sided pricing. This means the difficulty doesn't increase when you have two unknown values instead of one.

This matters because online marketplaces are everywhere, and small improvements in pricing can mean big money. For example, if a platform can learn optimal prices faster, it might offer better deals to buyers and sellers, leading to more trades and higher overall welfare. The result also raises a bigger question: does this 'no curse of dimensionality' hold for even more complex markets with many buyers and sellers?

While the algorithm is theoretical, it suggests that practical pricing tools could be built that learn efficiently from limited feedback. That could lead to smarter, more adaptive marketplaces that benefit everyone involved.