GRID-50: a pooled cross-sectional ranker on fifty coins

Abstract

Per-coin specialist models suffer a structural small-sample problem: one coin's history is one draw from one regime, and a promotion decision made on it is a bet on that regime persisting. A pooled cross-sectional ranker inverts the trade: one model, trained on every (coin, bar) cell of a 50-coin panel, learns only relative ordering — which names will outperform which over the next 24 hours — and is rewarded only when that ordering holds across the whole book. We expanded the tracked universe from 30 to 51 perpetuals specifically to give this model width, ran one pre-registered family through the full Grid A/B/C funnel, and report every number the funnel produced, including the one that embarrassed the construction sweep.

1 · Widening the universe: 30 → 51

Ranking breadth is the whole point of a cross-sectional model — K=8 longs and shorts out of 50 names is a meaningfully more selective book than out of 26. We screened every perpetual in the Hyperliquid archive for history depth and liquidity and added 21 names (XRP, WLD, DYDX, JTO, TAO, ADA, BCH, CRV, LTC, ZRO, TRX, DOT, FET, ONDO, kBONK, PYTH, HBAR, EIGEN, STRK, FIL, LDO), bringing the tracked set to 51, of which 50 are active (TON was delisted mid-program — its handling is its own small case study in Paper № 13). The data engineering behind the expansion — minute-level asset contexts from the exchange's S3 archive, hourly candles derived from snapshot mids where native candles do not reach — is documented in Paper № 10.

2 · The family, pre-registered

One signal recipe (OHLCV base, funding, open interest, cross-sectional momentum / funding / OI-velocity, realised volatility, price stress, calendar, BTC-lead features), one architecture family (LightGBM lambdarank), a small hyperparameter grid, horizons h=24 and h=48. The selection criterion — median-across-seeds pooled K=8/8 book Sharpe net 4 bps per leg — was written down before training started.

The coin-half gate deserves a note: it exists to catch a pooled model whose “edge” is actually three lucky names. The promoted family was positive on both halves of a random coin split for every one of five seeds — the edge is a property of the cross-section, not a handful of tickers.

3 · Grid C in full — the numbers that matter

All figures are the annualised Sharpe of the daily net return of a unit-gross K-of-N long/short book at 4 bps per leg on |Δw|, monthly expanding walk-forward June 2025 → June 2026, 394 non-overlapping daily bars. The seed distribution, not the best seed:

Note what is not claimed: the Grid B CPCV-panel Sharpe of +1.78 is a signal-detection number, not a tradeable one, and the causal +0.8–1.3 range is the honest expectation. CPCV fold models see chronologically later data than their test blocks; compression from +1.78 to this range on the causal panel was expected and is documented in the result file alongside per-seed diagnostics and plausibility flags (none fired).

4 · The finding that didn't cooperate

Grid B's construction sweep said: K=12, exit hysteresis, 3-day EMA smoothing — worth about +0.55 SR over a plain K=8 book, mostly by cutting turnover. On the causal walk-forward the ranking inverted: the plain K=8 book scored +1.30 against the smoothed K=12's +0.80. The mechanism is subtle but makes sense in hindsight — the CPCV panel averages 45 fold-models per bar, so its signal is already smooth, and construction-level smoothing tuned there over-dampens the noisier signal of a single causal expanding model.

The funnel's multiple-testing rules forbid a tie-breaking third sweep, so both books went to paper on 2026-07-03 — same predictions, different execution shape, adjudicated by live data. Whichever wins, the meta-lesson stands: execution parameters tuned on an ensembled research panel do not automatically transfer to the deployable object. We have not seen this failure mode named in the literature and suspect it is common in any shop that tunes constructions on cross-validated panels.

5 · Why pooled beats per-coin (for us, for now)

• Sample size. 50 coins × 2 years of daily cells is two orders of magnitude more supervision than any single coin provides, for one model selection decision instead of fifty.
• Market neutrality by construction. Long-K/short-K books cancel the directional beta that dominated our earlier specialist failures (Paper № 13).
• Selection resistance. One family, one pre-registered criterion, one promotion decision. The per-coin specialist grid made ~50 promotion decisions and paid for each in multiple-testing terms.
• The stacked alternative was debunked separately — per-coin models blended into a portfolio signal produced correlated drawdowns and no robust lift (Paper № 09).