Trade Execution Modeling

Idealized backtest: filled at the close, zero slippage
Real world:
1. The order book has a bid-ask spread
2. Large orders push prices (market impact)
3. Execution is delayed (there is latency from signal to fill)

No slippage model -> overly optimistic backtest -> losses in live trading

from src.quantlib.impact import fixed_slippage, linear_impact, sqrt_impact, delayed_execution

fixed_slippage(price=100.0, direction=1, bps=5.0)   # 100.05  (buy pays up)
fixed_slippage(price=100.0, direction=-1, bps=5.0)  #  99.95  (sell receives less)

# 100k shares against 1M ADV = 10% participation; at coeff 0.1 that is a 1% move.
linear_impact(price=100.0, direction=1, volume_traded=100_000, adv=1_000_000, impact_coeff=0.1)
# 101.0

# 250k against 1M ADV = 25% participation; 0.5 × 0.02 × sqrt(0.25) = 0.005 = 50bps.
sqrt_impact(price=100.0, direction=1, volume_traded=250_000, adv=1_000_000,
            volatility=0.02, eta=0.5)
# 100.5  (100.49999999999999 in binary floating point)

Backtest capital vs instrument ADV:
├── Capital < 0.5% of ADV -> fixed slippage (5bps) is enough
├── Capital 0.5-5% -> linear impact model
└── Capital > 5% -> square-root impact model (required)

Goal: execute at the day's volume-weighted average price

VWAP = Σ(Price_i × Volume_i) / Σ(Volume_i)

Execution logic:
1. Forecast the intraday volume profile (typically U-shaped)
2. Split the order according to the predicted profile
3. Execute proportionally in each time slice

Typical China A-share VWAP volume profile (U-shaped):
09:30-10:00  15%  (active open)
10:00-11:30  25%  (normal morning session)
13:00-14:00  15%  (weak afternoon session)
14:00-14:30  15%  (afternoon recovery)
14:30-15:00  30%  (active close)

VWAP in backtests:
- Daily backtest: use the VWAP field directly as the fill price
- Minute backtest: simulate VWAP order slicing

Goal: execute evenly over a specified time window

TWAP = simple time-sliced execution

Execution logic:
1. Define an execution window (for example 09:30-11:30)
2. Divide it into N time buckets
3. Execute total_size / N in each bucket

Pros and cons:
+ Simple, no need to forecast volume
- Easier to cause impact during low-volume periods
- Less adaptive than VWAP

signals = delayed_execution(raw_signal, delay_bars=1)   # T+1: trade tomorrow on today's signal
signals = delayed_execution(raw_signal, delay_bars=0)   # same-bar execution

Total trading cost = explicit cost + implicit cost

Explicit cost:
- Commission: China A-shares 2-3 bps, crypto 0.02-0.1%
- Stamp duty (China A-share sell side): 0.05% (sell orders only)
- Transfer fee: negligible

Implicit cost:
- Bid-ask spread: 0.5-5bps
- Market impact: depends on trade size and liquidity
- Opportunity cost: loss from not filling at the best price

{
  "commission": 0.001,
  "comment": "0.1% one-way commission, already includes stamp duty and spread"
}

{
  "commission": 0.001,
  "engine": "daily",
  "interval": "1D"
}

from src.quantlib.impact import delayed_execution


class SignalEngine:
    def __init__(self):
        # Execution assumption parameters
        self.execution_delay = 1       # T+1 delay
        self.slippage_bps = 5          # Fixed 5bps slippage
        self.max_participation = 0.05  # Maximum participation rate 5%

    def generate(self, data_map):
        for code, df in data_map.items():
            # 1. Generate raw signal
            raw_signal = self._compute_signal(df)

            # 2. Apply execution delay
            delayed_signal = delayed_execution(raw_signal, self.execution_delay)

            # 3. Apply volume filter (do not trade when liquidity is too low)
            volume_ok = df['volume'] > df['volume'].rolling(20).mean() * 0.3
            delayed_signal[~volume_ok] = 0

            signals[code] = delayed_signal

Step 1: Estimate annual turnover
  Annual turnover = annual trade count × 2 (buy + sell) / number of positions

Step 2: Compute annual cost drag
  Annual cost = annual turnover × total one-way cost

Step 3: Evaluate the impact on returns
  Net return = gross return - annual cost

Example:
  Annual turnover = 12 (monthly rebalance)
  One-way cost = 0.1%
  Annual cost = 12 × 0.1% = 1.2%
  If annualized return is only 5% -> costs eat 24% of returns!

### Backtest Results Under Different Slippage Assumptions

| Slippage (bps) | Annual Return | Sharpe | Max Drawdown |
|-----------|---------|--------|---------|
| 0 (ideal) | 15.2% | 1.35 | -18.5% |
| 3 | 13.8% | 1.22 | -19.0% |
| 5 | 12.9% | 1.15 | -19.2% |
| 10 | 11.1% | 0.98 | -19.8% |
| 20 | 7.5% | 0.65 | -20.5% |

Conclusion: the strategy still has meaningful profitability under 10bps slippage

## Execution Cost Analysis

### Strategy Trading Characteristics
| Metric | Value |
|------|-----|
| Average annual trade count | 48 |
| Annual turnover | 4.8x |
| Average holding days | 25 |
| Average order size | ¥50,000 |

### Cost Estimate
| Cost Item | Per Trade | Annualized |
|--------|------|------|
| Commission | 0.025% | 0.24% |
| Stamp duty | 0.025% | 0.12% |
| Estimated slippage | 0.03% | 0.29% |
| **Total** | **0.08%** | **0.65%** |

### Cost Impact
- Gross return: 12.5%
- Net return: 11.85%
- Cost drag: -0.65% (5.2% of gross return)
- Conclusion: cost impact is manageable

### Optimization Suggestions
1. Lower turnover (lengthen holding period)
2. Avoid trading during low-liquidity windows
3. Use limit orders instead of market orders