Architectures

Twenty-eight architectures, one shared metric.

All architecture cards below are ranked on the same number: best-cell walk-forward cross-sectional IC from the latest run. This replaces legacy references and keeps architecture comparison apples-to-apples.

28 empirically tested
11 families
116 hyperparameter cells
5 seeds per cell
ranking metric: best-cell WF IC
Analysis focus
Pooled cross-sectional ranker
Current focus: one LightGBM lambdarank family on a 50-coin panel — first to pass the full Grid A/B/C funnel; two K-of-N books in paper.
Single-coin specialists
Per-coin diagnostics and paper-trading microscope; retired as a live portfolio shape (debunk).
Architecture expansion
Next: run further families from the roster below through the same pre-registered funnel — survivors join the book pool.
Latest promotion · GRID-50 Grid A/B/C causal walk-forward · 2026-07-03 · live grid below is canonical
15 / 15
pre-registered seeds with positive causal walk-forward SR — LightGBM Rank, the first family through the full funnel
+0.8–1.3
seed-median causal SR range across book constructions (unit-gross K-of-N, 4 bps/leg, 13-month walk-forward)
50
coin universe of the pooled ranker (expanded from 26/30) — coin-half stability gate passed on every seed
2
K-of-N paper books trading side by side — the backtest could not separate their constructions, so paper adjudicates (Paper № 12)

Cards below are ranked by the shared screening metric (walk-forward cross-sectional IC from the Phase 1.7 grid, snapshot 2026-05-28). Screening rank is not deployment rank — the funnel survivor placed #16 on this screen; see architecture lessons. Use Training/Leaderboard for strategy-level analysis.