What transferred and what didn't: architecture lessons

1 · Benchmark rank is not universe rank

The controlled 918-experiment literature study places ModernTCN at the top of its crypto leaderboard; on our 26-coin Hyperliquid IC screen it ranked 24th of 28. The linear baselines that embarrass transformers on long-horizon benchmarks sat at the bottom of our table. Autoformer — mid-tier in the literature — was our #2 by screening IC. The pattern is not that the literature is wrong; it is that RMSE on long-horizon benchmarks and cross-sectional rank correlation on a noisy 24-hour horizon are different tasks, and transfer between them is close to random. The full ranking with methodology is in the architecture funnel paper.

2 · Twenty-eight architectures, roughly one signal

The more consequential finding was redundancy. Correlating per-coin IC profiles across the top arms: the TimeMixer / TSMixer / iTransformer cluster moves together at correlations above +0.95, with one pair (BITCN and StemGNN) at +0.999. Different mechanisms, same prediction. An ensemble of seven such arms is one signal counted seven times — the diversification lives in the taxonomy, not in the PnL. The genuinely independent arms were the unfashionable ones: rule-based cross-sectional deciles and the tabular LightGBM ranker.

3 · The screening winner is not the deployment winner

TimeMixer won the IC screen. The strongest standalone signals on the held-out strategy window came from a tuned GRU and the tabular ranker. And the family that eventually passed the full Grid A/B/C promotion funnel — fifteen out of fifteen seeds positive on a causal walk-forward (Paper № 12) — was LightGBM lambdarank, which had ranked a modest #16 on the original screen. Each stage of the pipeline asks a harder question than the last, and the answers reorder.

Why do trees keep beating sequence models here? Our working explanation, consistent with what the ablations show: at a 24-hour horizon on hourly bars, the predictive content is carried by a few dozen engineered, slowly-varying features — funding, open-interest dynamics, cross-sectional momentum, realised volatility. A gradient-boosted tree consumes those directly and spends its capacity on interactions. A sequence model must first rediscover the features from raw windows before it can use them, and on a few hundred thousand noisy panel rows it mostly doesn't. Sequence models earn their complexity on dense microstructure horizons — which is a different program (and needs the LOB tensor, not hourly candles).