acrl

almgren-chriss

Deep reinforcement learning for optimal execution of portfolio transactions.

Installation

pip install acrl

Usage

from collections import deque

import numpy as np

import acrl as sca
from acrl.agent import Agent

env = sca.MarketEnvironment()

agent = Agent(
    state_size=env.observation_space_dimension(),
    action_size=env.action_space_dimension(),
    random_seed=0,
)

liquidation_time = 60
n_trades = 60
risk_aversion = 1e-6
episodes = 10000

shortfall_hist = np.array([])
shortfall_deque = deque(maxlen=100)

for episode in range(episodes):
    current_state = env.reset(
        seed=episode,
        liquid_time=liquidation_time,
        num_trades=n_trades,
        lamb=risk_aversion,
    )

    env.start_transactions()

    for i in range(n_trades + 1):
        action = agent.act(current_state, add_noise=True)
        new_state, reward, done, info = env.step(action)
        agent.step(current_state, action, reward, new_state, done)
        current_state = new_state

        if info.done:
            shortfall_hist = np.append(shortfall_hist, info.implementation_shortfall)
            shortfall_deque.append(info.implementation_shortfall)
            break

    if (episode + 1) % 100 == 0:
        print(
            "\rEpisode [{}/{}]\tAverage Shortfall: ${:,.2f}".format(
                episode + 1, episodes, np.mean(shortfall_deque)
            )
        )

print("\nAverage Implementation Shortfall: ${:,.2f} \n".format(np.mean(shortfall_hist)))