Statsmaker: Probabilistic Programming Language for Market Microstructure Modeling

Statsmaker is a powerful Python library designed for market microstructure modeling, statistical analysis, and trading strategy development. It combines probabilistic programming, machine learning, and financial market microstructure theory to provide a comprehensive toolkit for researchers and traders.

Features

Probabilistic programming models based on Uber Pyro
Market microstructure models (order flow, limit order book, market making, etc.)
High-frequency trading and dark pool trading models
Liquidity, volatility, and execution quality metrics
Portfolio optimization and execution strategy optimization
Machine learning and reinforcement learning integration
Data simulation and market replay functionality
Price impact models including Almgren-Chriss, Kyle, Huberman-Stanzl, Bayesian Price Impact, and Bayesian Kyle models

Installation

Install statsmaker using pip:

pip install --upgrade statsmaker

Quick Start

Here are some examples demonstrating various features of statsmaker:

1. Order Flow Model

from statsmaker import StatsmakerBase, OrderFlowModel
import torch

# Create a Statsmaker instance
sm = StatsmakerBase()

# Define the order flow model
order_flow_model = OrderFlowModel()
sm.define_model("order_flow", order_flow_model.model)

# Prepare data
data = torch.randint(0, 2, (100,))

# Perform inference
inference_result = sm.infer("order_flow", data)

# Sample from the posterior distribution
posterior_samples = sm.sample("order_flow", inference_result.get_samples())

print(posterior_samples)

2. High-Frequency Trading Model

from statsmaker import HighFrequencyModel
import pandas as pd

# Prepare your market data
market_data = pd.DataFrame({
    'market_returns': [0.001, -0.002, 0.003, -0.001, 0.002],
    'volume': [1000, 1200, 800, 1100, 900]
})

# Create and fit the high-frequency model
hft_model = HighFrequencyModel()
fit_result = hft_model.fit(market_data)

# Make predictions
predictions = hft_model.predict(market_data, fit_result.get_samples())
print(predictions)

3. Portfolio Optimization

from statsmaker import MicrostructurePortfolioOptimizer
import numpy as np

# Prepare return data and microstructure data
returns = np.array([0.05, 0.03, 0.02, 0.04, 0.01])
microstructure_data = {
    'spread': [0.01, 0.015, 0.02, 0.01, 0.025],
    'volume': [10000, 8000, 12000, 9000, 11000]
}

# Create and use the portfolio optimizer
optimizer = MicrostructurePortfolioOptimizer(returns, microstructure_data)
optimal_weights = optimizer.optimize(risk_aversion=2)

print("Optimal portfolio weights:", optimal_weights)

4. Reinforcement Learning Trading Agent

from statsmaker import RLTrader
import pandas as pd

# Prepare your market data
market_data = pd.DataFrame({
    'price': [100, 101, 99, 102, 98, 103],
    'volume': [1000, 1200, 800, 1100, 900, 1300]
})

# Create and train the RL trader
rl_trader = RLTrader(market_data)
rl_trader.train(num_episodes=100)

# Make trading decisions
actions = rl_trader.act(market_data)
print("Actions:", actions)

5. Price Impact Models

Almgren-Chriss Model

from statsmaker import StatsmakerBase, AlmgrenChrissModel
sm = StatsmakerBase()
sm.define_model("almgren_chriss", AlmgrenChrissModel, sigma=0.02, gamma=0.1, eta=0.01)
impact = sm.calculate_impact("almgren_chriss", 1000, 100000)
print("Almgren-Chriss Price Impact:", impact)

Kyle Model

from statsmaker import StatsmakerBase, KyleModel
sm = StatsmakerBase()
sm.define_model("kyle", KyleModel, lambda_kyle=0.05)
impact = sm.calculate_impact("kyle", 1000)
print("Kyle Model Price Impact:", impact)

Huberman and Stanzl Model

from statsmaker import StatsmakerBase, HubermanStanzlModel
sm = StatsmakerBase()
sm.define_model("huberman_stanzl", HubermanStanzlModel, kappa=0.02, psi=0.1)
impact = sm.calculate_impact("huberman_stanzl", 1000, 100000)
print("Huberman-Stanzl Price Impact:", impact)

Bayesian Price Impact Model

from statsmaker import StatsmakerBase, BayesianPriceImpactModel
import pyro.distributions as dist
sm = StatsmakerBase()
alpha_prior = dist.Normal(0.0, 1.0)
beta_prior = dist.Normal(0.0, 1.0)
sm.define_model("bayesian_price_impact", BayesianPriceImpactModel, alpha_prior=alpha_prior, beta_prior=beta_prior)
order_sizes = [1000, 2000, 1500, 1200, 1800]
market_volumes = [100000, 150000, 120000, 110000, 130000]
price_impacts = [10, 20, 15, 12, 18]
params = sm.fit_model("bayesian_price_impact", order_sizes, market_volumes, price_impacts, num_steps=1000)
print("Fitted parameters (Bayesian Price Impact):", params)
predictions = sm.calculate_impact("bayesian_price_impact", order_sizes, market_volumes)
print("Predicted price impacts (Bayesian Price Impact):", predictions)

Bayesian Kyle Model

from statsmaker import StatsmakerBase, BayesianKyleModel
import pyro.distributions as dist
sm = StatsmakerBase()
lambda_prior = dist.Normal(0.0, 1.0)
sm.define_model("bayesian_kyle", BayesianKyleModel, lambda_prior=lambda_prior)
order_sizes_kyle = [1000, 2000, 1500, 1200, 1800]
price_impacts_kyle = [50, 100, 75, 60, 90]
lambda_kyle = sm.fit_model("bayesian_kyle", order_sizes_kyle, price_impacts_kyle, num_steps=1000)
print("Fitted lambda (Bayesian Kyle):", lambda_kyle)
predictions_kyle = sm.calculate_impact("bayesian_kyle", order_sizes_kyle)
print("Predicted price impacts (Bayesian Kyle):", predictions_kyle)

Contributing

We welcome contributions to statsmaker. If you have an idea for a new feature, a bug fix, or an improvement, please fork the repository and submit a pull request.

License

statsmaker is licensed under the Apache License, Version 2.0. See the LICENSE file for more details.

Acknowledgements

This project is built on the shoulders of giants. We want to acknowledge the contributions of the open-source community, particularly the developers of Pyro and other libraries that make probabilistic programming and financial modeling accessible.

jialuechen/statsmaker