Model white papers and methodology documentation
Multi-driver OLS regression on USD pairs at the 30-minute bar with 1-minute intra-bar stops. v3-revalidated, all four deflation gates pass, walk-forward Sharpe +7.94. 13 pairs (USDCNH excluded), all profitable. First Tier-1 strategy armed under the per-system attribution architecture.
FX residual against an equity-index driver via cumulative beta-adjusted residual + EMA crossover. v3 runs at hourly bars, 24/7, no regime gating — replacing the killed ER 15m 1-minute variant. Pending re-arm gated on the FV-Fast 24h clean window.
Donchian-style fade on hourly bars across 11 pairs (USDCNH excluded for pollution). The pre-v3 +2.08 Sharpe was a USDCNH-data artifact; clean re-run fails 5 of 8 strict deflation gates. Downgraded from Tier-1 admit to research-grade; live shadow continues observationally.
72 signal-emitting agents across three timescales feed into a two-tier architecture: Tier 1 standalone strategies (FV-Fast, ER, RT) trade their own signals, Tier 2 features gate or size them, Tier 3 is a meta-labeler — currently shadow-only after losing the 2026-05-11 A/B to HRP-Tier-1.
The three v3 Tier-1 sleeves (fv_fast_30m, er_1h, rt_1h) running on the canonical hourly :15 cadence against live market data. Ledger-only, no real orders. 30-day clean window + HRP-rolling warmup is the gate to live capital allocation (estimated mid-July 2026). 3-year backfill seeded for cross-machine determinism.
LightGBM meta-labeler v1 (frozen artifact meta_labeler_v1_2026_05_08, 21-agent panel). Lost the 2026-05-11 A/B vs HRP-Tier-1 by 3-8 Sharpe across two windows, so is not on live capital. Phase 2 retrain 2026-06-06 will include Tier-1 conviction streams as features — the canonical Lopez de Prado primary-mechanism-plus-meta-labeler role.
Pre-v3 regime-conditioning architecture (per-pair ATR + VIX / DXY / dispersion with Schmitt-trigger anti-flicker). The 30-40% Sharpe-lift claims did not survive the v3 deflation gauntlet under the clean substrate; regime gating is no longer live in any system. Retained as research history; state-encoded successors (vol_regime, cot_extremity) live on as feature-only signals.