THE RESEARCH PROCESS

310+ autonomous research sessions over two months. A hypothesis-driven methodology where dead ends are documented as rigorously as breakthroughs. What we chose not to deploy matters as much as what we did.

METHODOLOGY

01 HYPOTHESIS
→
02 EXPERIMENT
→
03 VALIDATION
→
04 DOCUMENT
→
05 DECISION

Every session follows the same five-step protocol. The fifth step is always a decision: deploy, iterate, or close as dead end. There is no ambiguity.

A journal entry is written before deep work begins, protecting findings if the session is interrupted. This is not a system that was built once and deployed. It was built, tested, rebuilt, tested again, and continually improved.

FIVE RESEARCH PHASES

PHASE 1: FOUNDATION
Core Infrastructure
Building core infrastructure: data pipeline, indicator library, backtest engine, regime detection, and risk management framework.
  • Dynamic stop-loss validation
  • Extreme value theory tail stops
  • Circuit breaker implementation
  • Optimal stopping theory integration
PHASE 2: THE BREAKTHROUGH PERIOD
Triple Gate Discovery
The period where the core intellectual property crystallized. Path-dependent targets and the triple gate framework emerged from a series of tightly connected experiments.
  • Re-entry lockout mechanism (saved 51% of losses)
  • Online adaptive convergence
  • Path-dependent targets breakthrough (MFE far more predictable than PnL, with MAE and MFE nearly orthogonal)
  • Triple gate architecture
  • Hybrid take-profit logic
PHASE 3: INTEGRATION AND STRESS TESTING
Ensemble Validation
Many individually promising ideas hurt when combined. This phase tested every component in concert and ruthlessly eliminated interactions that degraded ensemble performance.
  • Combinatorial purged cross-validation with zero probability of backtest overfitting
  • Information-theoretic hard gating proof
  • Online gradient boosting integration
  • Quad-gate architecture
  • Critical finding: individually effective components degraded performance when combined with the quad-gate ensemble
PHASE 4: MULTI-PAIR EXPANSION
Cross-Instrument Scaling
Scaling from single-pair to 5 currency pairs. Validating that the mathematical edge transfers across different equilibrium-seeking instruments.
  • First cross-pair deployment with strong out-of-sample performance
  • Primary backtest-to-live gap identified and root-caused
  • Gap resolution breakthrough, the single most impactful improvement for live performance
PHASE 5: PRODUCTION HARDENING
Live Deployment
Live trading deployment, real-time monitoring dashboard, forensic analysis tools, and continued pair expansion.
  • Forensic gate replay tooling
  • Multi-phase fleet deployment
  • Fleet monitoring and operational hardening

SESSION SPOTLIGHT: PATH-DEPENDENT TARGETS

THE LARGEST SINGLE IMPROVEMENT IN PROJECT HISTORY
“Predicting MFE separately from PnL should improve gating because MFE depends on signal strength while MAE depends on volatility.”
  • MFE proved substantially more predictable than direct PnL modeling
  • MAE-based gating produced the largest Sharpe ratio improvement of any single session
  • MAE and MFE were found to be nearly orthogonal, confirming independent information content
  • Path-dependent decomposition outperformed direct PnL gating by a wide margin
Led directly to the triple gate framework, the core risk architecture that filters every trade in the live system.

DEAD END DOCUMENTATION

Over 30% of research sessions resulted in dead ends, treated as valuable findings, not failures. Every dead end narrows the solution space and prevents future researchers from wasting time on paths already proven unproductive.

LINEAR FEATURE SATURATION
MULTIPLE CONFIRMATIONS
Advanced signal processing, information-theoretic measures, cross-pair analytics, microstructure proxies, and other orthogonal features all yielded zero predictive improvement. The linear feature space is saturated.
REGIME GATING
Regime-switching statistical models all fail for orderly declines. Regime detection identifies shifts after the damage is done, not before.
SOFT GATING
Hard gating dominates soft gating at extreme performance levels. Information-theoretic analysis confirms that the optimal gate converges to a binary decision under these conditions.
PROCESS MODEL CORRECTIONS
Alternative stochastic process specifications, all redundant or harmful when added to the existing framework.
WEIGHT OPTIMIZATION
Statistical shrinkage and robust optimization methods. Gate threshold discontinuity defeats smooth optimization methods.
EXIT PATCHING
Entry filtering subsumes exit patching. Blocking bad trades at entry is strictly superior to cutting them short after entry.

WHY DEAD ENDS MATTER

01
Demonstrate intellectual honesty. A research process that only reports successes is suspect.
02
Prevent future researchers from wasting time on paths already proven unproductive.
03
Define the boundaries of what is possible within the current framework.
04
The reasons failures fail reveal deeper truths about market structure.

RESEARCH INFRASTRUCTURE

310+
RESEARCH SESSIONS
11,760+
LINES OF JOURNAL
80+
ANALYSIS FILES
34
MATHEMATICAL INDICATOR SECTIONS
150+
DASHBOARD DEVELOPMENT SESSIONS
30%+
RESULTED IN DEAD ENDS

The ratio of documented dead ends to successful deployments is itself evidence of rigorous methodology. What was rejected matters as much as what was deployed.