# Ml4t Rl Execution

> Reinforcement learning for trade execution and hedging. Use when optimizing execution algorithms or dynamic hedging policies.

- Skill: `ml4t/ml4t-rl-execution` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ml4t/ml4t-rl-execution`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ml4t/ml4t-rl-execution/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: ml4t (https://skillmd.com/u/ml4t)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ml4t/ml4t-rl-execution

---

# RL for Trade Execution

Fixed execution schedules (TWAP, VWAP) ignore real-time market conditions. An RL agent adapts its execution rate based on order book state, reducing implementation shortfall.

## The Problem

Executing a large order at a fixed rate creates predictable market impact. A 100K-share TWAP sell ignores favorable liquidity bursts and pushes through thin books. The result: 20-50 bps of avoidable shortfall on institutional orders, compounding across thousands of trades per year.

## The Pattern

Model execution as a finite-horizon MDP. State: remaining shares, time left, volume, spread. Action: execution rate. Reward: negative implementation shortfall.

### WRONG

```python
# Static TWAP - ignores market conditions entirely
def twap_execute(total_shares: int, n_slices: int) -> list[int]:
    base = total_shares // n_slices
    remainder = total_shares % n_slices
    return [base + (1 if i < remainder else 0) for i in range(n_slices)]

schedule = twap_execute(100_000, 20)  # Same size every slice, blind to liquidity
```

### CORRECT

```python
import gymnasium as gym
import numpy as np
from gymnasium import spaces

class ExecutionEnv(gym.Env):
    """Agent decides what fraction of remaining shares to execute each step."""
    def __init__(self, total_shares=100_000, n_steps=20):
        super().__init__()
        self.total_shares, self.n_steps = total_shares, n_steps
        self.observation_space = spaces.Box(0, 1, shape=(4,), dtype=np.float32)
        self.action_space = spaces.Box(0, 1, shape=(1,), dtype=np.float32)

    def reset(self, seed=None, options=None):
        super().reset(seed=seed)
        self.remaining, self.step_idx = self.total_shares, 0
        self.arrival_price = 100.0
        return self._obs(), {}

    def step(self, action):
        shares = int(np.clip(action[0], 0, 1) * self.remaining)
        impact = 0.0001 * (shares / 5000)  # Linear market impact
        exec_price = self.arrival_price * (1 + impact)
        shortfall = (exec_price - self.arrival_price) / self.arrival_price
        reward = -abs(shortfall) * shares / self.total_shares
        self.remaining -= shares
        self.step_idx += 1
        done = self.step_idx >= self.n_steps or self.remaining <= 0
        if done and self.remaining > 0:
            reward -= 0.01  # Non-completion penalty
        return self._obs(), reward, done, False, {}

    def _obs(self):
        return np.array([
            self.remaining / self.total_shares, self.step_idx / self.n_steps,
            np.random.uniform(0.01, 0.05),  # spread
            np.random.uniform(0.3, 1.0),    # volume ratio
        ], dtype=np.float32)
```

## Guardrails

- **Non-completion penalty is mandatory** - without it the agent learns zero-trade is optimal
- **Normalize all state features** - raw share counts and prices break learning
- **Validate against TWAP baseline** - if RL underperforms TWAP, the environment is misconfigured
- **Use square-root impact for large orders** - linear impact underestimates cost at scale
- **Episode = one parent order** - do not mix multiple orders into one episode

## Production Implementation

`ml4t-backtest` provides execution simulation with realistic market impact:

```python
from ml4t.backtest import BacktestConfig, CommissionType, Engine
from ml4t.backtest.config import SlippageType
from ml4t.backtest.execution.impact import SquareRootImpact
from ml4t.backtest.execution.limits import VolumeParticipationLimit

config = BacktestConfig(
    commission_type=CommissionType.PER_SHARE,
    commission_per_share=0.005,
    slippage_type=SlippageType.VOLUME_BASED,
    slippage_rate=0.001,
)
env_engine = Engine(
    feed,
    strategy,
    config,
    market_impact_model=SquareRootImpact(volatility=0.02),
    execution_limits=VolumeParticipationLimit(max_participation=0.05),
)
```

## Checklist

- [ ] Environment has both time pressure and execution cost in the reward
- [ ] State is normalized (fractions, ratios) not raw values
- [ ] Non-completion is penalized (agent must finish the order)
- [ ] Trained agent beats TWAP baseline on test episodes
- [ ] Action space bounded (cannot execute more than remaining shares)