New Math Shows Why AI Struggles to Predict Data With Long Memories

Most AI training quietly assumes each piece of data is a fresh, independent event — like separate coin flips. But a lot of real-world data isn't like that at all. Financial markets, climate records, traffic patterns, electricity demand, and disease outbreaks all have what mathematicians call 'long memory': yesterday's numbers echo into today's, and last year's still echo too. In a new paper, statistician Elina Moldavskaya worked out exact mathematics for how AI models learn when the data behaves this way.

Her central result is a precise speed limit. When data has long memory, a model learns more slowly than it would on independent data — and she calculates exactly how much slower, based on how strong the memory is and how the model is structured. A second finding concerns 'weighting,' the common trick of telling a model to care more about recent data than old data. Weighting changes the fine details of how well the model performs, she shows, but it cannot change the underlying speed limit. In simple terms: you can fine-tune your approach, but you can't weight your way out of a fundamentally hard problem.

That matters because engineers already use this weighting trick constantly when building forecasting systems for energy grids, trading desks, and weather services. This paper gives the practice a rigorous foundation — and, just as usefully, tells practitioners where the ceiling is. If your data has strong long memory, no clever reweighting will make the model learn dramatically faster; you need either more data or a different modeling strategy.

The honest catch: this is pure mathematics. There is no code release, no product, and no immediate upgrade to anything you use. It also assumes a fairly idealized setup — smooth models trained on well-behaved data. Real-world data is messier, so the exact numbers may not transfer directly. Think of it as blueprints for the people who build forecasting tools, not a tool itself.