Each feature's SWA-averaged weight vs the latest live weight. Large differences = model still drifting at this feature.

The problem: Adam optimizer (and SGD) converges to sharp local minima in the loss landscape. Sharp minima often correspond to overfitted solutions — small changes in input cause big changes in prediction. They generalize poorly.

The fix (Izmailov et al. 2018): instead of using the final weights from training, average the weights from late training. The average sits in a flatter region of the loss landscape, where small input perturbations cause small output changes. Flatter minima have been shown to generalize better.

How it works here: every 60 seconds, the SWA module snapshots model.weights and updates a running average:
swa_w_new = (n × swa_w_old + new_w) / (n + 1) Once we have ≥5 snapshots, SWA.predict() uses the averaged weights for inference. The Unified Predictor blends SWA's prediction (14% weight) with the live model, multi-horizon ensemble, bootstrap, and k-NN.

Divergence as uncertainty signal: if the L2 distance between SWA and live model weights is large, the model is still drifting fast — high uncertainty. If divergence is small, the model has settled — lower uncertainty.