User asked if advanced architectures help. Tested ML vol-nowcaster (GB proxy) vs EMA for risk-layer
de-lever timing: GB OOS vol-corr +0.09 < EMA +0.18 (EMA captures clustering better); downstream
identical Sharpe +0.72, GB maxDD slightly worse. ML doesn't help -> CfC/Mamba2 won't either (more
overfit). Return-prediction already dead (IC 0.004); TLOB's order-book signal is latency-gated.
Fancy architectures add nothing; engine value = simple adaptive risk/allocation. Simple beats fancy,
confirmed once more. Adaptive multistrat book (EMA-vol, ~0.72 Sharpe, -8.5% maxDD) is the answer.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
