A new method called SGA tells you how much to trust an AI's multi-step forecast, not just what it predicted.
Researchers built a technique named Slicing-Graphing-Alignment (SGA) that measures uncertainty in time series foundation models - AI systems trained to predict sequences of future values, like demand or sensor readings, several steps ahead. The problem: these models don't just produce one forecast, they implicitly branch into a spreading tree of possible future paths, and some branches are far more accurate than others. SGA maps that branching structure as a directed graph and calculates how complex it is, combining the shape of the possible outcomes with the model's own randomness into a single uncertainty score. The team tested it against 11 different forecasting models and 27 datasets.
Knowing when to distrust a forecast matters more than the forecast itself in things like inventory planning or infrastructure monitoring, where acting on a bad multi-step prediction is expensive. SGA beat existing uncertainty methods at ranking which forecasts were likely to be wrong, and it surfaced a pattern worth noting: larger forecasting models produced more confident, lower-uncertainty predictions, hinting at another scaling law layered on top of the ones already tracked for accuracy.
Every foundation model runs into the confidence-interval problem eventually - time series models are just catching up to where language models already are.