Adam already adapts per-parameter step sizes. But the global learning rate is fixed at this.lr = 0.05. When the market regime shifts (concept drift), a higher LR helps the model adapt faster. When loss has settled, a lower LR enables fine-tuning.

The algorithm: every 5 minutes:

1. Read model.lossHistory (last 30 entries)
2. Fit linear regression: loss_t = a + b × t
3. If b > +0.001: loss is rising → drift detected → LR ×= 1.10
4. If b < -0.0008: loss is falling → converging → LR ×= 0.97
5. Otherwise: no change
6. Bound LR ∈ [0.01, 0.10]

The Adam optimizer's internal state (m, v moments) is preserved across LR changes, so the change is smooth — only the global step-size scaling changes.

Complements DriftPSI + AdversarialValidator: those modules detect drift; this module responds to it by making the model train faster.