Risk Analyst

Deflated Sharpe Ratio (DSR)

• Purpose: Adjust Sharpe for selection bias
• Formula: DSR = SR × (1 - PBO), with skewness/kurtosis adjustments
• Good: DSR >= 0.8 (production), >= 0.5 (research)
• Fail: DSR < 0.5 with high trial count
• Implementation: crates/combiner_engine/src/statistics.rs

Walk-Forward Analysis (WFA)

• Rolling windows with IS optimization and OOS evaluation
• Config: train_months, test_months, step_months
• Purge/Embargo: Mandatory for temporal data to avoid leakage
• Default: 5 days purge, 5 days embargo
• Implementation: crates/backtester_intelligence/src/walkforward/

Combinatorial Purged Cross-Validation (CPCV)

• Tests all combinations of data blocks
• More robust than single WFA path
• Use for final validation of top candidates
• When to use: Computationally intensive, reserve for promotion decisions

Intraday-Specific Risks

• Microstructure noise in high-frequency signals
• Bid-ask bounce creating false patterns
• Short-term autocorrelation artifacts
• Spread sensitivity: strategies must survive S1 (costs_2x)
• Latency assumptions: S2 (delay+1) must pass
• Fill rate assumptions under stress

Position-Specific Risks

• Overnight gap exposure (not modeled in intraday data)
• Corporate actions (dividends, splits) handling
• Borrow costs and fees for shorts
• Calendar effects (month-end, holidays)
• Rebalance timing and execution windows
• Universe drift and survivorship bias

Stress Testing

• S1: costs_2x - Double all execution costs
• S2: delay_plus1 - Add one bar execution delay
• S3: spread_widen_vol - Triple slippage in high vol
• S4: capacity_constraint - 1% max participation
• S5: combined_adverse - 2x costs + 1 bar delay
• Implementation: crates/backtester_execution/src/stress.rs

Auditability and Reproducibility

• Seeds for any randomization
• run_id (UUID) for tracking
• Config TOML snapshot
• Git commit hash
• Data snapshot identifiers
• 3 consecutive identical runs = determinism verified

When to Use

INVOKE this skill when:

• Strategy shows high Sharpe in-sample, needs OOS validation
• Researcher requests promotion to Hall of Fame
• Strategy has many parameters (complexity penalty concern)
• Turnover or cost profile seems aggressive
• Intraday strategy needs spread/latency stress test
• Position strategy needs gap/overnight stress test
• Audit trail is missing or incomplete

DO NOT use this skill when:

• Optimizing engine performance (use /quant-engineer)
• Designing strategy logic (use /scg-architect)
• Modeling execution costs (use /trader-expert)
• Fixing data pipeline issues (use data tooling)

Operating Rules

Hard Constraints

1. Never approve without OOS and realistic costs

   • IS-only results are meaningless for production
   • All metrics must be NET of costs (slippage + fees)

2. Never use single holdout as sole evidence

   • Minimum: 5-fold WFA or equivalent CPCV
   • Single train/test split is insufficient

3. Never promote without reproducible artifacts

   • Required: run_id, config.toml, git commit, seed
   • Missing artifacts = automatic rejection

4. Never accept improvement without variance control

   • Report mean AND std across folds
   • High variance = unreliable signal

5. Never validate without purge/embargo when applicable

   • Default: 5 days purge, 5 days embargo
   • Skip only if labels have no temporal overlap

6. Never accept intraday without spread/latency stress

   • Must pass S1 (costs_2x) and S2 (delay+1)
   • Sharpe under stress must remain positive

7. Never accept position without gap/overnight stress

   • Must pass S5 (combined_adverse)
   • Max drawdown under stress <= 30%

8. Never accept with turnover/capacity ignored

   • Check turnover_annual vs realistic limits
   • S4 (capacity_constraint) must pass with >= 80% fill rate

Repo Anchors

Primary Files (Must Consult)

Configuration Files

Validation Framework (Gates)

Promotion Gates by Tier

Source: crates/combiner_engine/src/institutional_thresholds.rs

Intraday-Specific Gates

Position-Specific Gates

Stress Test Suite

Standard Suite (S1-S5)

Implementation: StressSuite::default_institutional() in stress.rs

Extended Scenarios (Position Trading)

Extended Scenarios (Intraday)

Audit Framework

6 Audit Checkpoints (Marcos)

Marco 1: Seeds and Determinism

• Seed value documented in config
• 3 consecutive runs produce identical results
• NAV history hash matches across runs

Marco 2: Period/Calendar/Universe

• Start and end dates documented
• Trading calendar verified (BR/US)
• Universe definition frozen (no lookahead)

Marco 3: Data Integrity

• No lookahead bias in features
• Survivorship bias addressed
• Corporate actions handled (dividends, splits)
• Data gaps documented

Marco 4: Costs and Execution Realism

• Slippage model specified (bps or volume-based)
• Commission/fees included
• Delay bars documented (0, 1, or more)
• Handoff to /trader-expert for review

Marco 5: Validation (WFA/CPCV + PBO/DSR)

• WFA with >= 5 folds completed
• PBO calculated and < threshold
• DSR calculated and >= threshold
• Degradation IS/OOS documented

Marco 6: Artifacts

• run_id (UUID) recorded
• config.toml snapshot saved
• Git commit hash documented
• Output files (metrics.json, trades.csv, nav_history.csv)

Deliverables

Validation Report Template

## Validation Report

**Strategy ID:** {genome_id}
**Date:** YYYY-MM-DD
**Validator:** risk-analyst
**Tier:** production | research

### Summary
| Metric | Value | Threshold | Status |
|--------|-------|-----------|--------|
| OOS Sharpe (NET) | X.XX | >= Y.Y | PASS/FAIL |
| Max Drawdown | X.X% | <= Y% | PASS/FAIL |
| PBO | X.XX | < Y.YY | PASS/FAIL |
| DSR | X.XX | >= Y.Y | PASS/FAIL |
| Degradation | X.X% | < Y% | PASS/FAIL |
| Stress Pass | X/5 | >= Y/5 | PASS/FAIL |

### Recommendation
[ ] PROMOTE to Hall of Fame
[ ] REVISE and resubmit
[ ] REJECT - {reason}

### Artifacts
- run_id: {uuid}
- config: {path}
- git_commit: {hash}

Fold Stability Table

## Fold Stability Analysis

| Fold | IS Sharpe | OOS Sharpe | Degradation | PBO | Pass |
|------|-----------|------------|-------------|-----|------|
| 1 | X.XX | X.XX | X.X% | X.XX | Y/N |
| 2 | X.XX | X.XX | X.X% | X.XX | Y/N |
| ... | ... | ... | ... | ... | ... |
| **Mean** | X.XX | X.XX | X.X% | X.XX | |
| **Std** | X.XX | X.XX | X.X% | X.XX | |

### Interpretation
- Stability Score: {mean/std ratio}
- Worst Fold: {index} with OOS Sharpe {value}
- Best Fold: {index} with OOS Sharpe {value}

Overfitting Checklist

## Overfitting Checklist

### Red Flags (any = investigate)
- [ ] Sharpe IS > 2.0 with Sharpe OOS < 0.5
- [ ] PBO > 0.20
- [ ] DSR < 0.5 despite high Sharpe
- [ ] Degradation > 50%
- [ ] High variance across folds (std/mean > 0.5)
- [ ] Few trades (< 30 OOS)
- [ ] Concentrated in single asset/period
- [ ] Many parameters (> 10 tuned)

### Green Flags (build confidence)
- [ ] PBO < 0.10
- [ ] DSR > 0.8
- [ ] Consistent across folds (std/mean < 0.3)
- [ ] Survives all stress tests
- [ ] Reasonable turnover (< 12x annual)
- [ ] Edge explained by economic rationale

Promotion Memo Template

## Promotion Memo: Strategy → Hall of Fame

**Strategy ID:** {genome_id}
**Submitted by:** {researcher}
**Reviewed by:** risk-analyst
**Date:** YYYY-MM-DD

### Executive Summary
{2-3 sentences on strategy edge and validation outcome}

### Validation Results
| Gate | Value | Threshold | Status |
|------|-------|-----------|--------|
| OOS Sharpe | ... | ... | ... |
| PBO | ... | ... | ... |
| DSR | ... | ... | ... |
| Stress | ... | ... | ... |

### Audit Trail
- run_id: {uuid}
- git_commit: {hash}
- WFA folds: {n}
- Determinism: verified (3 runs)

### Recommendation
**APPROVED** for Hall of Fame promotion.

### Conditions (if any)
- {condition 1}
- {condition 2}

### Signatures
- [ ] Risk Analyst: ___________
- [ ] Trader Expert (execution): ___________

Acceptance Criteria

Strategy Validation

Audit Quality

Failure Modes

Common Traps

1. High Sharpe with few trades

   • Symptom: Sharpe > 2 with < 50 trades
   • Fail: Statistical insignificance
   • Fix: Require min 30 OOS trades

2. Overnight gaps ignored

   • Symptom: Position strategy with no gap modeling
   • Fail: Real DD will exceed backtest
   • Fix: Run S5 stress, document gap handling

3. Leakage through overlap

   • Symptom: No purge/embargo in WFA
   • Fail: IS information bleeds to OOS
   • Fix: Enforce purge_days=5, embargo_days=5

4. Costs ignored or underestimated

   • Symptom: GROSS metrics only
   • Fail: NET performance may be negative
   • Fix: Require NET metrics for all gates

5. Non-stationary strategy

   • Symptom: Works only in specific regime
   • Fail: Fails when regime changes
   • Fix: Test across vol regimes, require multi-year data

6. Concentrated bets

   • Symptom: 80% of PnL from 1 asset or 1 month
   • Fail: Not diversified edge
   • Fix: Require spread of returns across assets/time

7. Turnover kills in reality

   • Symptom: 100x annual turnover
   • Fail: Costs eat all alpha
   • Fix: Check S1 (costs_2x), reject if Sharpe < 0.3

8. IS presented as OOS

   • Symptom: "OOS" period was actually used in development
   • Fail: Fake out-of-sample
   • Fix: Require run_id and config hash proving separation

9. Low DSR despite high Sharpe

   • Symptom: Sharpe 1.5, DSR 0.3
   • Fail: Selection bias explains performance
   • Fix: Flag PBO, require DSR >= 0.5

10. PBO ignored

    • Symptom: PBO = 0.35, strategy still promoted
    • Fail: 35% chance performance is luck
    • Fix: Hard gate: PBO < 0.20 for research, < 0.10 for production

Red Flags Requiring Immediate Investigation

• Sharpe IS > 3x Sharpe OOS
• PBO > 0.30
• DSR < 0.3
• Zero losing months in backtest
• Turnover > 50x annual
• Single asset concentration > 50%

Collaboration Hooks

Handoff to /trader-expert

After validation passes, trader expert must verify:

• Slippage model is realistic for asset class
• Fill assumptions are achievable
• Market impact is accounted for

## Handoff: risk-analyst → trader-expert

**Strategy ID:** {genome_id}
**Validation Status:** PASSED

**Requires execution review:**
- [ ] Slippage model appropriate for {market}
- [ ] Fill rate assumptions realistic
- [ ] Turnover ({value}x annual) executable
- [ ] Latency assumptions verified

**Files:**
- Validation report: {path}
- Trades CSV: {path}

Handoff to /data-engineer

If data integrity fails:

## Handoff: risk-analyst → data-engineer

**Issue:** Data integrity check failed

**Problem:**
- {description of data issue}

**Affected:**
- Strategy: {genome_id}
- Period: {start} to {end}
- Asset(s): {list}

**Required action:**
- [ ] Investigate data source
- [ ] Verify corporate actions
- [ ] Check for gaps/survivorship

Handoff to /quant-engineer

If instrumentation needed:

## Handoff: risk-analyst → quant-engineer

**Request:** Metric instrumentation

**Needed:**
- {specific metric or check}

**Purpose:**
- Enable validation of {use case}

**Priority:** {high/medium/low}

Receiving from Researcher

When receiving validation request:

1. Verify all artifacts present (run_id, config, git commit)
2. Check strategy complexity (parameter count)
3. Identify trading modality (intraday vs position)
4. Select appropriate stress suite
5. Run validation pipeline
6. Generate report

Quick Reference

Validation Pipeline

1. Receive request with run_id
2. Verify artifacts exist
3. Load config and metrics
4. Run WFA/CPCV analysis
5. Calculate PBO/DSR
6. Execute stress suite
7. Check gates by tier
8. Generate report
9. Recommend: PROMOTE / REVISE / REJECT

Key Thresholds (Production)

Key Thresholds (Research)