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.
GRID-50 update (July 2026): the funnel survivor
LightGBM lambdarank is the first — and so far only — family to pass the full Grid A/B/C promotion funnel end-to-end: CPCV signal gates including a random coin-half stability split, an execution-cost sweep on locked predictions, and a 15-seed causal walk-forward with 15/15 seeds positive (seed-median SR +0.8 to +1.3 at 4 bps per leg, unit-gross K-of-N book). It is now the signal engine of the two GRID-50 market-neutral paper books on a 50-coin universe — the full write-up covers the run, including the construction-transfer surprise.
The instructive part: this family ranked a mid-table #16 of 28 on the generalist IC screen below. Screening metrics rank architectures by average discrimination; deployment gates rank them by robustness, cost survival, and seed stability — and the orderings differ (architecture lessons).
Pros and cons on this universe
Pros
- First family through the full Grid A/B/C funnel — 15/15 seeds positive on the causal walk-forward (GRID-50, 2026-07).
- Distinctness score 0.61 vs the modern mixer cluster — by far our most independent neural-adjacent arm.
- Lambdarank natively suited to cross-sectional rank tasks.
- Fast training, fast inference, no GPU required — 15-seed retrains are cheap.
Cons / failure modes
- Mid-rank IC (+0.113) on the generalist screen — was misreported +0.146 before audit.
- Earlier single-strategy deployment (STRAT-04b) decayed live and was retired — see the backtest→paper→live decomposition.
- Cross-sectional rank target evaluation is easy to substitute for raw-return Sharpe (see IC is not Sharpe).