Below is a curated list of articles of interest for machine learning based algorithmic trading. Eventually we will provide code samples that will implement a few of these articles. Feel free to email us if you have any suggested papers.
Trading algorithms using Reinforcement Learning
- Deep Hedging [https://arxiv.org/abs/1802.03042]
- Provides a framework for hedging a portfolio of derivatives in the presence of market frictions such as transaction costs, market impact, liquidity constraints or risk limits using modern deep reinforcement machine learning methods.
- Implementation would be somewhat complex although the paper introduces some neat convex risk measures. It also attempts to compare strategies based on different level of risk and network complexity.
- Date published: February 2018
- Architecture: Recurrent neural network, with 2 fully-connected hidden layers
- Time horizon: 30 days for training, daily rebalancing
- Input data: undisclosed price time series and other trading signals for equity derivatives
- Reward function: Convex risk measure, with consideration for transaction costs and payoffs
- Universe: Undisclosed – assumed to be American and European options
Trading algorithms using Machine Learning
- Using Macroeconomic Forecasts to Improve Mean Reverting Trading Strategies [https://arxiv.org/abs/1705.08022]
- Provides a pair trading strategy on a basket of currencies that is biased by an SVM model on macroeconomic data. The initial pairs trading strategy is derived by modelling a standard Ornstein-Uhlenbeck process.
- Date published: May 2017
- Architecture: SVM – details undisclosed
- Time horizon: 1995-2008 for training, 2008-2014 for testing
- Input data: undisclosed price time series (on currencies) and other macroeconomic indicators (S&P 500 Index, Federal Funds Rate, 10-Year Treasury)
- Universe: 7 currencies paired with USD (AUDUSD, EURUSD, GBPUSD, NZDUSD, CADUSD, CHFUSD, JPYUSD)
Trading algorithms using Deep Learning
- Deep Stock Representation Learning: From Candlestick Charts to Investment Decisions [https://arxiv.org/abs/1709.03803]
- Provides a portfolio construction strategy based on clustering of stocks on deep features of resulting graphs. Stock selection based on Sharpe ratio.
- The actual idea is neat, although it would be seem that learning from a graphical representation, instead of from the raw data itself, would be an information-lossy transformation.
- Date published: February 2018
- Architecture: Convolutional AutoEncoder
- Input data: 4-channel price time series (open, high, low, close) for equity
- Time horizon: 20 days for training, holding period of 10 days
- Intermediate data: Candlestick chart image
- Universe: FTSE 100
- Continual Learning Augmented Investment Decisions: [https://arxiv.org/abs/1812.02340]
- Provides a framework for a long/short investment strategy where continual learning augmentation is used. Context is defined as similar market environments under a dynamic time warping similarity measure.
- Certainly the first paper in finance with continual learning augmentation. Neat idea.
- Date published: December 2018
- Architecture: Feed forward neural network (undisclosed details). Dynamic time warping for market similarity.
- Input data: Stock level characteristics using style factor excess returns
- Time horizon: 6 months training, 6 months rebalancing
- Universe: 4500 equities over 30 countries – All countries World Ex-USA Equities Index between 2001 and 2017
- A Machine Learning Framework for Stock Selection [https://arxiv.org/abs/1806.01743]