hjnormey/DOLPHIN
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
master
DOLPHIN/nautilus_dolphin/Registry.md
hjnormey 01c19662cb initial: import DOLPHIN baseline 2026-04-21 from dolphinng5_predict working tree 
Includes core prod + GREEN/BLUE subsystems:
- prod/ (BLUE harness, configs, scripts, docs)
- nautilus_dolphin/ (GREEN Nautilus-native impl + dvae/ preserved)
- adaptive_exit/ (AEM engine + models/bucket_assignments.pkl)
- Observability/ (EsoF advisor, TUI, dashboards)
- external_factors/ (EsoF producer)
- mc_forewarning_qlabs_fork/ (MC regime/envelope)

Excludes runtime caches, logs, backups, and reproducible artifacts per .gitignore.
2026-04-21 16:58:38 +02:00

5.4 KiB
Executable File
Raw Permalink Blame History

DOLPHIN NG — Performance Registry

Canonical benchmark tiers. Update this file whenever a new result becomes the production target. GOLD = what the system must beat or match. SILVER = previous gold. BRONZE = regression floor.

🥇 D_LIQ_GOLD — Active Production Candidate (2026-03-15)

Metric Value vs prev GOLD vs BRONZE
ROI 181.81% +85.26 pp +93.26 pp
DD 17.65% +3.33 pp +2.60 pp
Calmar 10.30 vs 6.74 vs 5.88
Trades 2155 identical identical
avg_leverage 4.09x — —
liquidation_stops 1 (0.05%) — —

Engine: LiquidationGuardEngine(soft=8x, hard=9x, mc_ref=5x, margin_buffer=0.95, adaptive_beta=True) Factory: create_d_liq_engine(**engine_kwargs) — also create_boost_engine() default Module: nautilus_dolphin/nautilus/proxy_boost_engine.py Config key: engine.boost_mode = "d_liq" (now DEFAULT_BOOST_MODE)

Mechanism:

  • Inherits adaptive_beta scale_boost from AdaptiveBoostEngine (GOLD)
  • Leverage ceiling raised to 8x soft / 9x hard (from 5x/6x)
  • MC-Forewarner assessed at 5x reference (decoupled) → 0 RED/ORANGE/halted days
  • Liquidation floor stop at 10.6% adverse move (= 1/9 × 0.95) — prevents exchange force-close
  • DD plateau: each +1x above 7x costs only +0.12pp DD (vs +2.6pp for 5→6x)

Validation (exp9b, 2026-03-15):

  • All 4 leverage configs compared vs unguarded (exp9): B/C/D all improved ROI + reduced DD
  • E (9/10x): 5 liquidation stops → cascade → dead; D (8/9x) is the sweet spot
  • pytest -m slow tests/test_proxy_boost_production.py → 9/9 PASSED (2026-03-15)
  • MC completely silent: 0 RED, 0 ORANGE, 0 halted across 56 days at 8/9x
  • Trade count identical to Silver (2155) — no entry/exit timing change

Compounding ($25k, 56-day periods):

Periods ~Time Value
3 ~5 mo $559,493
6 ~1 yr $12,521,315
12 ~2 yr $6,271,333,381

🥈 GOLD (prev) — Former Production (demoted 2026-03-15)

Metric Value
ROI 96.55%
DD 14.32%
Calmar 6.74
Trades 2155
scale_mean 1.088
alpha_eff_mean 1.429

Engine: AdaptiveBoostEngine(threshold=0.35, alpha=1.0, adaptive_beta=True) Factory: create_boost_engine(mode='adaptive_beta') — non-default, opt-in for conservative/quiet-regime use Validation: pytest -m slow tests/test_proxy_boost_production.py → 7/7 PASSED 2026-03-15

🥉 BRONZE — Regression Floor (former silver, 2026-03-15)

Metric Value
ROI 88.55%
PF 1.215
DD 15.05%
Sharpe 4.38
Trades 2155

Engine: NDAlphaEngine (no proxy_B boost) Equivalent factory call: create_boost_engine(mode='none', ...) Validation script: test_pf_dynamic_beta_validate.py

Bronze is the absolute regression floor. Falling below Bronze on both ROI and DD is a failure.

All Boost Modes (exp8 results, 2026-03-14)

mode ROI% DD% ΔDD ΔROI Notes
none (Bronze) 88.55 15.05 — — Baseline
fixed 93.61 14.51 −0.54 +5.06 thr=0.35, a=1.0
adaptive_alpha 93.40 14.51 −0.54 +4.86 alpha×boost
adaptive_thr 94.13 14.51 −0.54 +5.58 thr÷boost
adaptive_both 94.11 14.51 −0.54 +5.57 both combined
adaptive_beta ⭐ 96.55 14.32 −0.72 +8.00 alpha×(1+day_beta) — prev GOLD

Extended Leverage Configs (exp9b results, 2026-03-15)

Config ROI% DD% Calmar liq_stops Notes
GOLD (5/6x) 96.55 14.32 6.74 0 adaptive_beta baseline
B_liq (6/7x) 124.01 15.97 7.77 1 improved vs unguarded
C_liq (7/8x) 155.60 17.18 9.05 1 improved vs unguarded
D_liq (8/9x) 181.81 17.65 10.30 1 D_LIQ_GOLD
E_liq (9/10x) 155.88 31.79 4.90 5 cascade — dead

Test Suite

# Fast unit tests only (no data needed, ~5 seconds)
pytest tests/test_proxy_boost_production.py -m "not slow" -v

# Full e2e regression (55-day backtests, ~60 minutes)
pytest tests/test_proxy_boost_production.py -m slow -v

Unit tests: ~40 (factory, engine, extended leverage, liquidation guard, actor import) E2E tests: 9 (baseline + 5 boost modes + winner-beats-baseline + D_liq repro + MC silent)

Last full run: 2026-03-15 — 9/9 PASSED, exit code 0 (50:20)

Promotion Checklist

To promote a new result to D_LIQ_GOLD (production):

  1. Beats prev GOLD on ROI (+85pp); DD increased +3.33pp but Calmar +53% — acceptable
  2. Trade count identical (2155) — no re-entry cascade
  3. MC completely silent at mc_ref=5.0 — 0 RED/ORANGE/halted
  4. liquidation_stops=1 (0.05%) — negligible, no cascade
  5. pytest -m slow passes — 9/9 PASSED (2026-03-15, 50:20)
  6. Updated Registry.md, memory/benchmarks.md, memory/MEMORY.md
  7. create_d_liq_engine() and classes added to proxy_boost_engine.py
  8. Wire create_d_liq_engine into DolphinActor as configurable option