Old weather records hold surprises: the thermometer suddenly jumps by half a degree not because of the weather, but because the instrument was replaced at the station. Such sharp shifts in data are called change points, and separating them from true climate signals is a problem statisticians grapple with. Conventional methods either see too many jumps or miss important ones.

A restorer of an ancient fresco removes layer after layer, and with each pass their hand becomes more precise at distinguishing genuine cracks from random scratches. The new IRFL method works the same way: it makes several passes over the data, each time penalizing small changes that look like noise while preserving large, real ones. Instead of cutting out everything that seems uneven in one go, the algorithm gradually filters out false signals.

On data from Mauna Loa volcano, where carbon dioxide was measured for 768 consecutive months, IRFL found five change points. They coincided with real events: the Pinatubo eruption and El Niño episodes. The method also removes noise from images while preserving sharp edges, which helps doctors and agronomists. A restorer, with each pass, more accurately separates genuine cracks from random scratches. IRFL does the same with data: it filters out false jumps and leaves only those changes that truly altered the picture.