Scientists Found a Better Ruler for Measuring How Predictable Your Movements Are

Every time you pull up a ride-share app, check a delivery time, or see traffic alerts, an AI is trying to guess where you and thousands of others will go next. But until now, there was no reliable way to know how close those guesses come to the absolute best possible prediction. This paper introduces BER-PEF, a new measuring stick built on a statistical concept called the Bayes error rate — that's the mathematical limit of how accurate any prediction can ever be.

The core problem the researchers solved is that the "true" predictability of a person's location history is invisible. You can watch someone travel for years and never know if your algorithm found every pattern. So they designed a framework that converts location data — such as GPS tracks, check-in sequences, and contextual details like time of day — into a common format. Then it tests how far different prediction methods stray from that theoretical limit, on a scale from "no better than guessing" to "perfectly predictable." It's like grading students not by raw scores but by how much room they left for improvement, even when you don't know the perfect score.

In tests on real mobility datasets — including Foursquare check-ins in New York City and Tokyo, GPS tracks from Chinese taxis, and detailed personal trajectories — the team showed that their method gave fairer scores than existing toolkits. Their estimates also matched real-world performance when data was deliberately scrambled, which proved the system wasn't fooling itself. Because the approach works across different types of location data and even different input formats, it becomes the first common benchmark for the entire field.

Why should a non-tech person care? The practical payoff comes in a few years: location-based services will be tuned with honest comparisons, so your navigation, package tracking, and ride-share ETA could become noticeably more reliable. Meanwhile, scientists and product developers will finally have a universal way to say "this model is doing better than any other we've tested" or "this approach has hit its ceiling — time to try a different method."