A new paper says your weather app's forecasting model might be arguing with itself behind the scenes.
Researchers built AliO (Align Outputs), a method for long-term time series forecasting models - the systems that predict things like electricity demand or weather weeks out using a sliding window of past data. They found that when you feed these models overlapping, time-shifted versions of the same input, the models often produce different predictions for the exact same future timestamp, a consistency problem that standard error metrics like mean squared error don't even measure. To catch it, the team built a new metric called TAM (Time Alignment Metric), which checks whether a model agrees with itself across overlapping forecasts rather than just checking how close it lands to the truth. Applying AliO to existing models cut that self-contradiction by up to 58.2% and, in a bonus result, improved raw forecasting accuracy by up to 27.5% in some tests.
Most forecasting benchmarks reward being close to the right answer on average, not being internally consistent. A model that predicts a 30% chance of rain for next Tuesday one day and 60% the next, using almost the same data, can still score well on MSE while being useless for anyone trying to plan around it. That gap between benchmark performance and real-world trustworthiness is exactly what keeps forecasting tools out of decisions that matter.
The fix also doubles as a quiet admission: the field has spent years optimizing for a number that was never measuring the thing people actually care about.