Operating Model
Overview
The Edge Factory operates as a continuous pipeline: Research → Validate → Deploy → Monitor → Retire.
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ RESEARCH │───▶│ VALIDATE │───▶│ DEPLOY │───▶│ MONITOR │───▶│ RETIRE │
│ (Ideas) │ │ (Backtest) │ │ (Paper→Live)│ │ (Track) │ │ (Sunset) │
└─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘
│ │ │ │ │
▼ ▼ ▼ ▼ ▼
Hypothesis Pass/Fail Staged Drift Alert Archive +
Generation Gates Rollout Response Post-mortem
Stage Gates
Gate 1: Research → Validation
| Criterion | Pass | Fail |
| Hypothesis documented | Written rationale exists | No documentation |
| Data requirements defined | Sources + frequency specified | Unclear data needs |
| Initial signal logic | Pseudocode or prototype | Vague concept only |
Gate 2: Validation → Deployment
| Criterion | Pass | Fail |
| In-sample Sharpe | ≥ 1.2 | < 1.2 |
| Out-of-sample Sharpe | ≥ 0.8 | < 0.8 |
| OOS degradation | < 30% | ≥ 30% |
| Walk-forward consistency | 60%+ windows profitable | < 60% |
| Max drawdown (backtest) | < 8% | ≥ 8% |
| Anti-bias audit | All tests pass | Any test fails |
| Independent review | Approved | Rejected |
Gate 3: Deployment Stages
| Stage | Duration | Criteria to Advance |
| Paper trading | 2-4 weeks min | Matches backtest within 20% |
| Live (minimal size) | 2-4 weeks | No execution issues, stable P&L |
| Live (target size) | Ongoing | Continuous monitoring |
Gate 4: Monitoring → Retire
| Trigger | Action |
| Drawdown > 50% of max allowed | Pause strategy, investigate |
| 30-day Sharpe < 0.5 | Flag for review |
| 60-day Sharpe < 0.3 | Mandatory review meeting |
| 90-day cumulative loss | Retire strategy |
| Regime shift detected | Reduce size, investigate |
Roles and Responsibilities
| Role | Responsibility | Current Assignment |
| Strategy Researcher | Generate hypotheses, prototype signals | Founder + Claude agents |
| Quant Engineer | Implement backtests, feature engineering | Founder + Claude agents |
| Risk Manager | Set limits, monitor exposure, approve deployments | Founder |
| Auditor | Independent validation, adversarial testing | Claude audit agent |
| Operator | Execute deployments, incident response | Founder |
| Function | Tool | Notes |
| Strategy development | Pine Script v6 | TradingView indicators/strategies |
| Backtesting | Python (vectorbt/backtrader) | Local execution |
| Live signals | TradingView alerts → Webhook | JSON format |
| Execution (FX) | MT5 | Via MQL5 or API bridge |
| Execution (Crypto) | Exchange APIs | Binance, Bybit, etc. |
| Paper trading | TradingView Paper | Or broker demo accounts |
| Monitoring | Custom dashboards | Metrics collection TBD |
| Version control | Git | All code and configs |
| Documentation | Markdown in repo | Single source of truth |
Workflow Cadence
| Activity | Frequency | Owner |
| Strategy review | Weekly | Founder |
| Risk limit check | Daily (automated) | Monitoring system |
| Backtest refresh | Monthly or on data update | Quant pipeline |
| Full audit | Per strategy, pre-deployment | Audit agent |
| Post-mortem (retired strategies) | Within 1 week of retirement | Founder |
Communication
- Alerts: Webhook → Discord/Telegram for trade signals
- Incidents: Logged in
docs/runbooks/incident_log.md - Decisions: Logged in
research/reports/ with rationale
Constraints
- No manual overrides in live trading: System trades as programmed or not at all
- All changes via version control: No direct edits to live configs
- Staged rollouts only: Never deploy directly to full live size
- Document before deploy: No undocumented strategies go live