Mechanism diagram · Tabular Tree
How it works
Tabular features per coin per bar (funding-rate z-score, OI velocity, momentum, volatility, order-book pressure, etc.) are fed to LightGBM.
The lambdarank objective optimises the relative cross-sectional rank at each bar, not per-coin regression error.
Output scores are used as cross-sectional rankings for the strategy.
Pros and cons on this universe
Pros
- Distinctness score 0.61 vs the modern mixer cluster — by far our most independent neural-adjacent arm.
- Deployed in STRAT-04b with +1.51 walk-forward Sharpe at 4 bps per leg.
- Lambdarank natively suited to cross-sectional rank tasks.
- Fast training, fast inference, no GPU required.
Cons / failure modes
- Mid-rank IC (+0.113) — was misreported +0.146 before audit.
- Pass A is fragile (−4.24 in STRAT-04b) — regime-conditional model with no built-in regime gate.
- Cross-sectional rank target evaluation is easy to substitute for raw-return Sharpe (see Paper № 03).