Understanding the superiority of multi-model ensemble forecasts through reservoir computing

What the math gives to ML

The paper identifies a concrete mechanism behind multi-model ensemble improvement in chaotic forecasting: independently trained reservoir models produce forecast errors that often lie along the target trajectory's unstable manifold, with different models deviating in opposite directions. If these errors are approximately independent and unbiased, arithmetic averaging reduces the error magnitude as N^{-1/2}, but heavy-tailed error distributions create deviations from this law. The most transferable neural-network construction is an instability-aware ensemble forecaster that estimates the local unstable direction, measures signed forecast errors along it, and adaptively chooses between ordinary averaging and tail-robust aggregation while monitoring the predicted ensemble-size scaling.