sidmohan0/quant-train

Python Repository to ingest, feature engineer, train, backtest, and run a random forest model to predict the direction of the S&P500 at the start of the next day's trading session.

quant-train

Overview

S&P 500 Overnight Random Forest Systematic Trading Strategy with data ingestion using Polygon API, feature engineering, storage into a MySQL DB, a backtesting framework with visualizations, and code for a production deployment.

I'll be updating this regularly to add more functionality. Take a peek at the Future Updates section for more details. Interested in collaborating, contributing, or just want to chat? Please reach out to me direct at [email protected].

Overivew

This repo is a collection of python scripts that do the following:

• download SPY spot and options data.
• calculate overnight returns and volatility
• solve for the greeks of the option contracts (check out feature_functions.py)
• train a random forest model to predict overnight returns
• backtest the model
• calculate the performance metrics of the model
• send an email with the prediction using gmail

Installation

1. Clone the repo
2. Use a virtual env , I use Pyenv w/ Python 3.8.10 running on Ubuntu 20.04
3. pip install -r requirements.txt

Future Updates

Quality of Life

• pipenv or poetry
• refactor feature engineering parts
• refactor backtesting parts
• refactor into packages and modules

Data

• Use the polygon python client more proactively to speed up some of the ingestion of raw information like options contracts

Automation

• run_backtest.sh
• run_prod.sh
• some sort of cron job

Containerization

• Dockerfile
• Orchestrator (Kubernetes, Docker Swarm, etc.) and deployment scripts - also potentially making Automation section redundant

Processing

By far the slowest part of the process is the feature engineering. While we are ultimately dumping a lot of information about contracts in-memory when running the program, the feature engineering, where we calculate the Greeks of the contracts, involves various operations over the dataframes. Some things I'm thinking about and open to suggestions on:

• Polars for faster processing of Dataframes
• GPU acceleration
• PySpark for distributed processing, also use w/ or instead of Polars
• Maybe add and or update the options functions in feature_functions.py to use tf-quant-finance

Modeling

• Hyperparameter tuning
• Feature selection framework

Acknowledgement:

Credit to QuantGalore for the original code. I've dressed it up a bit but the core pieces as of initial commit i owe to them. Check out their repo, and blog post

Changelog

11/3/23

• Added .env
• Added requirements.txt
• Refactoring, file renaming
• Abstracted out MySQL connection
• Source: SP500 Overnight Prediction