backtest-expert — скилл для Claude Code

---
name: backtest-expert
description: Expert guidance for systematic backtesting of trading strategies. Use when developing, testing, stress-testing, or validating quantitative trading strategies. Covers "beating ideas to death" methodology, parameter robustness testing, slippage modeling, bias prevention, and interpreting backtest results. Applicable when user asks about backtesting, strategy validation, robustness testing, avoiding overfitting, or systematic trading development.
---

# Backtest Expert

Systematic approach to backtesting trading strategies based on professional methodology that prioritizes robustness over optimistic results.

## Core Philosophy

**Goal**: Find strategies that "break the least", not strategies that "profit the most" on paper.

**Principle**: Add friction, stress test assumptions, and see what survives. If a strategy holds up under pessimistic conditions, it's more likely to work in live trading.

## When to Use This Skill

Use this skill when:
- Developing or validating systematic trading strategies
- Evaluating whether a trading idea is robust enough for live implementation
- Troubleshooting why a backtest might be misleading
- Learning proper backtesting methodology
- Avoiding common pitfalls (curve-fitting, look-ahead bias, survivorship bias)
- Assessing parameter sensitivity and regime dependence
- Setting realistic expectations for slippage and execution costs

## Prerequisites

- Python 3.9+ (for evaluation script)
- No API keys required
- No external data dependencies — metrics are user-provided

## Workflow

### 1. State the Hypothesis

Define the edge in one sentence.

**Example**: "Stocks that gap up >3% on earnings and pull back to previous day's close within first hour provide mean-reversion opportunity."

If you can't articulate the edge clearly, don't proceed to testing.

### 2. Codify Rules with Zero Discretion

Define with complete specificity:
- **Entry**: Exact conditions, timing, price type
- **Exit**: Stop loss, profit target, time-based exit
- **Position sizing**: Fixed $$, % of portfolio, volatility-adjusted
- **Filters**: Market cap, volume, sector, volatility conditions
- **Universe**: What instruments are eligible

**Critical**: No subjective judgment allowed. Every decision must be rule-based and unambiguous.

### 3. Run Initial Backtest

Test over:
- **Minimum 5 years** (preferably 10+)
- **Multiple market regimes** (bull, bear, high/low volatility)
- **Realistic costs**: Commissions + conservative slippage

Examine initial results for basic viability. If fundamentally broken, iterate on hypothesis.

### 4. Stress Test the Strategy

This is where 80% of testing time should be spent.

**Parameter sensitivity**:
- Test stop loss at 50%, 75%, 100%, 125%, 150% of baseline
- Test profit target at 80%, 90%, 100%, 110%, 120% of baseline
- Vary entry/exit timing by ±15-30 minutes
- Look for "plateaus" of stable performance, not narrow spikes

**Execution friction**:
- Increase slippage to 1.5-2x typical estimates
- Model worst-case fills (buy at ask+1 tick, sell at bid-1 tick)
- Add realistic order rejection scenarios
- Test with pessimistic commission structures

**Time robustness**:
- Analyze year-by-year performance
- Require positive expectancy in majority of years
- Ensure strategy doesn't rely on 1-2 exceptional periods
- Test in different market regimes separately

**Sample size**:
- Absolute minimum: 30 trades
- Preferred: 100+ trades
- High confidence: 200+ trades

### 5. Out-of-Sample Validation

**Walk-forward analysis**:
1. Optimize on training period (e.g., Year 1-3)
2. Test on validation period (Year 4)
3. Roll forward and repeat
4. Compare in-sample vs out-of-sample performance

**Warning signs**:
- Out-of-sample <50% of in-sample performance
- Need frequent parameter re-optimization
- Parameters change dramatically between periods

### 6. Evaluate Results

**Questions to answer**:
- Does edge survive pessimistic assumptions?
- Is performance stable across parameter variations?
- Does strategy work in multiple market regimes?
- Is sample size sufficient for statistical confidence?
- Are results realistic, not "too good to be true"?

**Decision criteria**:
- ✅ **Deploy**: Survives all stress tests with acceptable performance
- 🔄 **Refine**: Core logic sound but needs parameter adjustment
- ❌ **Abandon**: Fails stress tests or relies on fragile assumptions

Use the evaluation script for a structured, quantitative assessment:

```bash
python3 skills/backtest-expert/scripts/evaluate_backtest.py \
  --total-trades 150 \
  --win-rate 62 \
  --avg-win-pct 1.8 \
  --avg-loss-pct 1.2 \
  --max-drawdown-pct 15 \
  --years-tested 8 \
  --num-parameters 3 \
  --slippage-tested \
  --output-dir reports/
```

The script scores across 5 dimensions (Sample Size, Expectancy, Risk Management, Robustness, Execution Realism). Treat the machine-readable `decision` and `blocking_reasons` as authoritative. The 0–100 `quality_score` is diagnostic: even a high score cannot override non-positive expectancy (except that fewer than 30 trades yields `NOT_EVALUABLE` first), a 50% or greater drawdown, a stricter user-supplied drawdown limit, insufficient trades, missing slippage/friction validation, a test period shorter than five years, or results that look too good (win rate above 90% with max drawdown below 5%, reason `too_good`, which needs a look-ahead and data audit before adoption). `verdict` remains a conservative Deploy/Refine/Abandon compatibility field that is never better than the score-based verdict. Use `--max-acceptable-drawdown-pct` to set a personal limit below the hard 50% ceiling; equality with the personal limit is allowed. Invalid inputs exit with status 1 and no report.

## Key Testing Principles

### Punish the Strategy

Add friction everywhere:
- Commissions higher than reality
- Slippage 1.5-2x typical
- Worst-case fills
- Order rejections
- Partial fills

**Rationale**: Strategies that survive pessimistic assumptions often outperform in live trading.

### Seek Plateaus, Not Peaks

Look for parameter ranges where performance is stable, not optimal values that create performance spikes.

**Good**: Strategy profitable with stop loss anywhere from 1.5% to 3.0%
**Bad**: Strategy only works with stop loss at exactly 2.13%

Stable performance indicates genuine edge; narrow optima suggest curve-fitting.

### Test All Cases, Not Cherry-Picked Examples

**Wrong approach**: Study hand-picked "market leaders" that worked
**Right approach**: Test every stock that met criteria, including those that failed

Selective examples create survivorship bias and overestimate strategy quality.

### Separate Idea Generation from Validation

**Intuition**: Useful for generating hypotheses
**Validation**: Must be purely data-driven

Never let attachment to an idea influence interpretation of test results.

## Common Failure Patterns

Recognize these patterns early to save time:

1. **Parameter sensitivity**: Only works with exact parameter values
2. **Regime-specific**: Great in some years, terrible in others
3. **Slippage sensitivity**: Unprofitable when realistic costs added
4. **Small sample**: Too few trades for statistical confidence
5. **Look-ahead bias**: "Too good to be true" results
6. **Over-optimization**: Many parameters, poor out-of-sample results

See `references/failed_tests.md` for detailed examples and diagnostic framework.

## Output

- `reports/backtest_eval_<timestamp>.json` — structured evaluation with per-dimension scores, red flags, `decision`, `blocking_reasons`, and verdict
- `reports/backtest_eval_<timestamp>.md` — human-readable report with dimension table, key metrics, and red flag details

Key JSON fields:

- `decision`: authoritative result, one of `DEPLOY`, `REFINE`, `ABANDON`, `REJECT`, `RISK_LIMIT_EXCEEDED`, `VALIDATION_REQUIRED`, `NOT_EVALUABLE`. Gate precedence: `NOT_EVALUABLE` (fewer than 30 trades or profit-factor overflow), then `REJECT`, then `RISK_LIMIT_EXCEEDED`, then `VALIDATION_REQUIRED` (no slippage test, under 5 years, or `too_good`). Only when no gate blocks does the score yield `DEPLOY`/`REFINE`/`ABANDON`.
- `blocking_reasons`: IDs of every gate that fired (`profit_factor_overflow`, `small_sample`, `negative_expectancy`, `zero_expectancy`, `excessive_drawdown`, `risk_limit_exceeded`, `no_slippage_test`, `short_test_period`, `too_good`); all are preserved even when a higher-precedence gate sets the decision.
- `profit_factor`: may be `null`. `profit_factor_status` explains why: `FINITE` (numeric value), `NO_LOSSES` (positive gross profit, no gross loss), `UNDEFINED_ZERO_GROSS` (no gross profit and no gross loss), or `OVERFLOW` (ratio overflow despite nonzero losses).

## Resources

### Methodology Reference
**File**: `references/methodology.md`

**When to read**: For detailed guidance on specific testing techniques.

**Contents**:
- Stress testing methods
- Parameter sensitivity analysis
- Slippage and friction modeling
- Sample size requirements
- Market regime classification
- Common biases and pitfalls (survivorship, look-ahead, curve-fitting, etc.)

### Failed Tests Reference
**File**: `references/failed_tests.md`

**When to read**: When strategy fails tests, or learning from past mistakes.

**Contents**:
- Why failures are valuable
- Common failure patterns with examples
- Case study documentation framework
- Red flags checklist for evaluating backtests

## Critical Reminders

**Time allocation**: Spend 20% generating ideas, 80% trying to break them.

**Context-free requirement**: If strategy requires "perfect context" to work, it's not robust enough for systematic trading.

**Red flag**: If backtest results look too good (>90% win rate, minimal drawdowns, perfect timing), audit carefully for look-ahead bias or data issues.

**Tool limitations**: Understand your backtesting platform's quirks (interpolation methods, handling of low liquidity, data alignment issues).

**Statistical significance**: Small edges require large sample sizes to prove. 5% edge per trade needs 100+ trades to distinguish from luck.

## Discretionary vs Systematic Differences

This skill focuses on **systematic/quantitative** backtesting where:
- All rules are codified in advance
- No discretion or "feel" in execution
- Testing happens on all historical examples, not cherry-picked cases
- Context (news, macro) is deliberately stripped out

Discretionary traders study differently—this skill may not apply to setups requiring subjective judgment.