Skip to main content Link Menu Expand (external link) Document Search Copy Copied
Algorithmic Trading System

Status Updates

About Us

If you don’t know us already, our names are Jin Kim and Sai Chintalapati, both seniors in SEAS studying Computer Science.

Project Idea

The goal of our project is to develop a high-performance system to buy and sell securities for profit, specifically cryptocurrencies, with the decisions being made by a deep learning model. This way, the analysis and trading is automated, similar to what is done at today’s hedge funds. The trading of cryptocurrency was chosen over stocks due to trading being consistent around the clock, rather than fixed trading hours. The infrastructure should allow for the execution of trades on a mock exchange API that trades cryptocurrencies in a “paper-trading” account. The success of the algorithmic trading system can be measured by the return of the system historically, and in a specific time-frame.

Wins

  • End-to-end system online running on Google Cloud, including infrastructure for data pipeline, triggering algorithm every 20 minutes, and running trades
  • Accompanying web dashboard (http://stocktradingbotdashboard.uw.r.appspot.com/)
  • Rapidly brainstormed and iterated through various trading engine algorithm ideas, from simple moving average to mean reversion to momentum trading strategies
  • Created an ensemble model of these three strategies, optimizing the parameters for their weights. Backtesting: 75% in profits over the course of two yeras, despite the turmoil in sudden increase and decrease of prices in the time period tested, 1/1/2021 to 2/28/2023

Plans

  • We believe a system ingesting stock data in real-time and making real-time decisions is a more exciting thing to work on, so we plan to transition our current system to that method
  • Coded portfolio rebalncing with Sharpe Ratio, but yet to use because better option may be available with https://pypi.org/project/pyportfolioopt/
  • Create a backtest infrastucture, for now we’ve been using QuantConnect

Issues

There were some issues we faced along the way, primarily regarding the success of the underlying stock prediction algorithms. From experimenting with various short-term (intraday) price prediction models (using LSTM, random forest, and SVMs), we were unable to create a model that produced reliable hourly predictions that were in any way usable (48% accuracy for guessing what the price movement will be in the next hour). We are unblocked at the moment, building out other parts of the bot and validating that the returns are replicable using backtesting.