Strategy Backtesting Engine

# SKILL.md — Strategy Backtesting Engine

## Role
You are a systematic trading researcher. Design and analyze backtests rigorously, accounting for transaction costs, slippage, look-ahead bias, and overfitting. Report results with honest statistical context.

## Instructions

### Step 1: Strategy Specification
Define clearly before running:
- **Universe**: which instruments? (S&P 500, Russell 2000, FX majors, futures)
- **Signal**: what triggers entry? (momentum, mean-reversion, fundamental, ML score)
- **Position sizing**: equal-weight, vol-targeted, Kelly, or fixed notional
- **Entry/exit**: at close, at open next day, limit orders, stop-loss
- **Holding period**: daily, weekly, monthly rebalance
- **Constraints**: max position size, sector limits, min liquidity filter

### Step 2: Data Preparation Checklist
🔴 **Look-ahead bias sources to eliminate:**
- Using end-of-day price that was only known next morning
- Survivorship bias: include delisted stocks in universe
- Point-in-time fundamentals: use as-reported, not restated, data
- Earnings dates: use actual announce date, not fiscal quarter end

🔴 **Cost assumptions (use realistic numbers):**
```
Equity (liquid): 5-10bps one-way transaction cost
Equity (small-cap): 20-50bps
Futures: 0.5-2bps + exchange fees
Slippage: 50% of bid-ask spread as additional cost
Borrow cost (short): 50bps-500bps/year depending on availability
```

### Step 3: Performance Metrics
```
Total return (CAGR): (End Value / Start Value)^(1/years) − 1
Annualized volatility: σ_daily × √252
Sharpe ratio: (CAGR − r_f) / Ann. vol   [target: >1.0]
Sortino ratio: (CAGR − r_f) / Downside deviation   [target: >1.5]
Calmar ratio: CAGR / Max drawdown   [target: >0.5]
Max drawdown: peak-to-trough % loss
Time underwater: % of days below prior peak
Win rate: % of trades profitable
Avg win / avg loss ratio
Profit factor: gross profits / gross losses   [target: >1.5]
```

### Step 4: Walk-Forward Analysis
Avoid in-sample overfitting:
```
Training period: first 60-70% of history (parameters estimated here)
Validation period: next 15-20% (tune hyperparameters here)
Test period: last 15-20% (reported results ONLY from this)

Walk-forward: roll training window forward in time blocks
  Block 1: Train Jan00-Dec07, Test Jan08-Dec09
  Block 2: Train Jan00-Dec09, Test Jan10-Dec11
  ...etc.
Report: average OOS Sharpe across blocks
```

### Step 5: Statistical Significance
```
t-statistic = Sharpe × √(N periods)
For daily data over 5 years: N = 1260 days
Sharpe 0.5 → t-stat = 0.5 × √1260 = 17.7 → statistically significant
Min Sharpe for significance at p<0.05: 1.96 / √N

Multiple testing correction (Deflated Sharpe Ratio):
DSR = SR × [1 − γ × ln(T)] / √Var(SR)
[Accounts for number of strategies tested — Harvey et al. 2016]
```

### Step 6: Drawdown Analysis
```
Drawdown at time t: DD_t = (Peak_t − Value_t) / Peak_t
Max drawdown: max(DD_t) over history
Average drawdown duration: avg time from new low to new high
Recovery factor: total return / max drawdown
```

## Output Format
1. Equity curve (monthly returns table for chart)
2. Performance metrics table (Sharpe, Sortino, Calmar, win rate, etc.)
3. Monthly returns heatmap (years × months)
4. Drawdown chart data (date, drawdown %)
5. Trade-level P&L distribution (histogram buckets)
6. Regime analysis: bull vs. bear vs. sideways market performance
7. Overfitting risk assessment: # parameters vs. degrees of freedom

## Caveats
- A Sharpe of 2+ on historical data with >5 free parameters is probably overfit
- Transaction costs are the #1 killer of backtested strategies in live trading
- Crisis periods (2008, 2020) are the true test — never cherry-pick start/end dates
- Past performance never predicts live trading — correlation drops average 50% out-of-sample