Where the ensemble’s Sharpe concentrates by market regime, and the nine steps that built the stack. Dated 2026-09-27; regime row is the run #11 pooled benchmark, both generator families.
Ensemble Sharpe by regime
OVERALL row of the run #11 heatmap, pooled standalone net Sharpe:
| Regime | trend_up | trend_down | range | vol_shock | whale_dump | whale_pump | capitulation | blowoff_top | accumulation | distribution |
|---|---|---|---|---|---|---|---|---|---|---|
| Ensemble Sharpe | 2.138 | −2.717 | −1.556 | 0.573 | 6.872 | 5.429 | 2.927 | 12.472 | 1.424 | −1.840 |
- Alpha concentrates where on-chain flow leads price:
whale_dump,whale_pump,blowoff_top. trend_downanddistributionare drains the book absorbs rather than shorts;rangeis hedged bybollinger_fade@0.25.- Caveat: standalone unscaled Sharpes; per-cell standard error ≈ ±2 (±4–6 for the sparse
vol_shock/capitulationslices) — extreme cells are not point forecasts.
Layering recipe
Reproduce in this order (IN median = median path net Sharpe on the benchmark fixtures):
| # | Step | What moved | IN med after |
|---|---|---|---|
| 1 | Sharpe-weight the catalog, clip negatives (run #2) | −1.734 → +1.258 | +1.258 |
| 2 | Nonnegative coordinate descent on IN+HO, objective mean + 0.25·worst (run #5) | re-derived weights; IN-only tuning shown to overfit (IN 2.32 / HO 0.83) | +1.647 |
| 3 | Turnover taming — discrete stateful deadbands, not output smoothing (run #6) | turnover −72 % whale_tx_spike, −62 % stablecoin, −19 % shock_cont, −8 % macd; output-EMA smoothing rejected (gross 0.89 → 0.86) | +1.872 |
| 4 | Episodic rewrites + mutes — whale_flow_composite state machine with funding-overheat veto (run #7), descent re-run (run #8) | +0.89, then +0.37 | +3.131 |
| 5 | Inverse-vol sizing on the combined position (run #9) | biggest single step (+1.08); 27-cell grid chose win 24 h / target 0.02 / cap 5 (plateau 12–48 / 0.02–0.03 / 3–5) | +4.213 |
| 6 | Surgical event signal — whale_pump_frontrun @3 + catalog hygiene (run #10) | +0.31 | +4.524 |
| 7 | Hygiene pass — dev_commit_trend v2 @5 (weekly aggregate, conditional decay short), sopr_regime v2 @0 (run #11) | +0.06 | +4.585 |
| 8 | Segment-2 whale stack — false-lead-in resistance: composite commits at half size until 6 h momentum confirms, frontrun gains a fast lane + bounded episodes, whale_tx_spike sizes half until the 3-bar price move confirms the strike; oi_funding_squeeze catalogued @0 (run #13) | +0.31 pooled | 3.134 |
| 9 | Fat-tail shock gating — shock_continuation v2 needs positive, elevated netflow on 2 of 3 bars, skips IV-exhausted and isolated no-whale tails, bounded 12-bar episode; keltner 0.25 → 0 (run #14) | +0.18 pooled | 3.314 |
Segment-1 rows score the family-1 fixtures only; segment-2 rows are pooled medians over both families (10 paths).
Stack shape
1h reactive → 4h swing → 1d trend → shared weighted mix → inverse-vol sizing (24 h window, 0.02 target, 5× cap) → position clip. This is a reading order, not a serial gating pipeline: signals compute independently on their declared frames, completed higher-timeframe values carry forward causally, and no timeframe can veto another. The ensemble normalizes by total active weight, then scales by the inverse-vol governor and clips to [−1, +1].
Funded members (9, weight sum 22.0): whale_tx_spike 5.0 · dev_commit_trend 5.0 · whale_flow_composite 3.0 · whale_pump_frontrun 3.0 · shock_continuation 3.0 · stablecoin_supply_growth 2.0 · macd_hist 0.5 · bollinger_fade 0.25 · adx_trend_strength 0.25.