A new ensemble method quietly fixes one of forecasting's oldest problems: letting a single bad model wreck the average.
Researchers built a system called Conformal Adversarial Generative Ensemble, or CAGE. It runs several generative models to produce competing forecasts, then uses a separate discriminator to score each one with a statistical technique called conformal prediction. That scoring produces a p-value for each forecast, which the system uses to down-weight predictions it doesn't trust. In tests on New Zealand milk collection data and global monkeypox case counts, CAGE beat standard ensemble methods, particularly when the data included outliers or noise.
Most forecasting ensembles just average predictions, which means one wildly wrong model can drag down an otherwise solid group. CAGE's conformal prediction step gives it a mathematical way to say "this forecast looks unreliable" before it ever reaches the final number. That matters for anyone forecasting demand, disease spread, or weather, where a single garbage data point can cascade into a bad decision.
The method has only been tested on two fairly niche datasets so far, so whether it holds up on messier real-world finance or supply-chain data is still an open question.