An interpreter at a meeting doesn't memorize every word the speaker says. They pick out the key ideas and then convey the condensed meaning without losing the details. That's roughly how the new model for predicting El Niño works. Ocean temperature is recorded as a three-dimensional table: longitude, latitude, and depth. Each slice contains thousands of numbers, and spotting patterns directly in them is nearly impossible.

The model first decomposes the data into a small number of factors, or "themes," that describe typical temperature patterns. It then looks for how these factors in the past relate to factors in the future. The relationship turns out to be simple and compact: instead of millions of coefficients, only a few hundred need to be estimated. But pure compression isn't enough. The ocean has local anomalies, such as a sudden warming off the coast of Peru. To handle these, the model includes a sparse component that catches rare but important events that don't fit the overall pattern.

The spatial maps of the factors matched known El Niño regions. This means the method doesn't just produce accurate numbers; it also shows where the future climate shift is actually taking shape. Now this approach can be applied to other data where time and space are intertwined, from brain activity to financial markets.