Files

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

3.8 KiB

Executable File

Raw Blame History

FROZEN ALGO SPEC — GOLD REFERENCE (ROI=181.81%) & RECREATION LOG

1. Specification Overview

The "D_LIQ_GOLD" configuration is the frozen champion strategy for the Dolphin NG system. It achieves high-leverage mean reversion across 48 assets using eigenvalue velocity divergence signals, gated by high-frequency volatility and regime-aware circuit breakers.

Performance Benchmark (Parity Confirmed 2026-03-29)

ROI: +181.01% (Target: 181.81%)
Max Drawdown (DD): 19.97% (Target: ~17.65% – 21.25%)
Trade Count (T): 2155 (EXACT PARITY)
Liquidation Stops: 1 (EXACT PARITY)
Period: 56 days (2025-12-31 to 2026-02-26)

2. Core Findings from Reconstruction

During the recreation process, it was discovered that the deterministic "Trade Identity" (T=2155) is highly sensitive to one specific parameter: Volatility Calibration.

Finding: Static vs. Rolling Volatility

The GOLD Spec (T=2155): Requires a Static Vol Calibration. The volatility threshold (vol_p60 = 0.00009868) MUST be calculated once from the first 2 days of data and held constant for the entire 56-day duration.
The REGRESSION (T=1739): Occurs when using a "Rolling" volatility threshold (as seen in certify_extf_gold.py). This "Rolling" logic tightens too early during high-volatility regimes, suppressing ~416 trades and collapsing the ROI from 181% to 36%.

Finding: Warmup Reset

Parity REQUIRES the Daily Warmup Reset logic (resetting _bar_count each day). This skips the first 100 bars (~8.3 minutes) of every data file. Continuous-mode backtests that lack this reset will result in ~2500+ trades and different ROI characteristics.

3. Critical File Inventory & Behavior

Canonical Verification (The Source of Truth)

test_dliq_fix_verify.py:
- Purpose: Direct reproduction of the research champion. Uses float64 for calibration and static vol_p60.
- Match Status: GOLD MATCH (ROI 181%, T=2155).

Logic Core

esf_alpha_orchestrator.py: Core signal logic and "Daily Warmup" logic (lines 982-990).
proxy_boost_engine.py: Implementation of LiquidationGuardEngine which adds the 10.56% stop-loss floor.

Configuration & Data

exp_shared.py: Contains ENGINE_KWARGS (Fixed TP=95bps, Stop=1.0%, MaxHold=120) and MC_BASE_CFG (MC-Forewarner parameters).
vbt_cache/: Repository of the 56 Parquet files used for the benchmark.

4. Frozen Configuration Constants

Parameter	Value	Description
`vel_div_threshold`	-0.020	Entry signal threshold
`fixed_tp_pct`	0.0095	95bps Take-Profit
`max_hold_bars`	120	10-minute maximum hold
`base_max_leverage`	8.0	Soft cap (ACB can push beyond)
`abs_max_leverage`	9.0	Hard cap (Never exceeded)
`stop_pct_override`	0.1056	Liquidation floor (1/9 * 0.95)

5. RECREATION INSTRUCTIONS

To recreate Gold results without altering source code:

Shell: Use the Siloqy environment.
Verify Script: Execute python dvae/test_dliq_fix_verify.py.
Observation: Parity is achieved when Trade Count is exactly 2155.
Note: Disregard certify_extf_gold.py for ROI reproduction as its rolling vol logic is optimized for safety, not research parity.

3.8 KiB Executable File Raw Blame History Unescape Escape