Table of Contents
1 JP Morgan RL in Financial Markets
- financial markets are difficult for RL because of market impact
1.1 Action Space
- Small action spaces are easy to learn and large are difficult,
- To small of an action space, however, and the agent wont be adaquately enabled for its market
1.2 Testing
- simulation? How do you simulate the financial markets? How can you trust that the model will accurately capture the markets?
- What metric to use for offline performance evaluation?
1.3 Rewards in RL
- Naive objective would be to increase profit
- but how does that account for underdelivering? market impact?
2 Two Sigma - Forecasting Financial Returns using ML
2.1 Speakers
- Dai Shi
- Xiang Zhou
2.2 Efficent market hypothesis
- Two Sigma claims market not at efficent as it may seem
- Opportunies in the irrationalities of the market
2.3 Identifying opportunity
- Using ML to find mispriced securities
- Can we predict equity returns?
2.3.1 Example
- Data source: instagram
- metric: product popularity
- predict: current and future price of the stock
- Use mutliple models for multiple tasks
2.4 Workflow
- ideation, data, forecast research, simulation, release
2.5 Build a model from historical data that will forcast unseen future
- (out of sample, test) select a period in historical data and pretend its unseen future
- (in sample, train) Hide and research the stategy with the rest of the historical data
- Start with a hypothesis
- For example: Product Popularity -> Stock Returns
- Make the hypothesis testable
- Postive sentiment incomments -> positive future 90 day returns