Date: January 21, 2024

Topic: Machine Learning at Hedge Funds

Recall

Model output based on collected data

Notes

Gist of Machine Learning

Given some data, we create a model that with an input $x$ we can predict an output $y$

• Possible $x$ values: Bollinger value, price momentum, current price
• Possible $y$ values: Future price, future return

Given some historical $x, y$ data, can we predict for future $x$ what $y$ will be?

Supervised Regression Learning

Make a numerical prediction based on example input and actual outputs

Examples

• Linear Regression (parametric)
  • Data is thrown away after the best regression model is found
• K Nearest Neighbour (instance based)
  • Data is kept and new data is evaluated to see how it best fits existing data
• Decision Trees
  • $x$ goes down a tree structure until a terminal node (leaf) gives the prediction
• Decision Forests
  • Collection of decision trees

Stocks have natural labels, where we can just look forward in the market to get our labelled data.

By having a bunch of features $x$ (Bollinger value, current price etc), we can train an ML model to predict prices

Using with Stock Data

• For example, if we use a look forward of 5 days, we get a pairing of input features $x$ and output price $y$
• This allows us to have labelled data, since we just need to look forward to get the actual output for training

Example use-case

To test algorithms, we can roll back time and run in on past data.

From this, we can make orders in the past and forecast our portfolio earnings to compare with with the actual earnings

Backtesting

Roll back time and test your system with your model

A previous order was made at t=t1, and another order made at t=t2

• By rolling back time, we can test our algorithm on our portfolio
• The performance can then be measured by metrics like Sharpe Ratio

By backtesting, we can see how our strategy performs against benchmarks (S&P 500 in this case)

Regression based learning has a number of issues associated

<aside> 📌 SUMMARY: We can use supervised regression learning to optimize our portfolio, and perform backtesting to see how well it does against the market

</aside>