Data Lessons from Algorithmic Trading (part 3): "Think In Terms of Experiments"

Posted by Q McCallum on 2020-05-20

This article is part of a series. In Part 1, I outlined the premise: ML/AI shops can borrow tips and best practices from how algorithmic trading (“algo trading”) shops operate. The rest of the articles explore those ideas in more detail.

The stock market is not entirely predictable, so even with the best of research behind a trading strategy, things may not turn out as planned. 1 Traders constantly develop and test new strategies in order to see what works best given the current (and, expected future) market conditions.

Every trading strategy is, at its core, an experiment.

They key for any algo trading shop, then, is to get very good at designing, building, executing, and learning from those experiments. Algo trading shops want to know: “What if we buy according to this pattern?” “What if we take this approach when the unemployment numbers come out?” To answer such questions, their researchers pull historical market data, develop ideas, test those ideas in a simulated environment, and then put them out into the real stock market to buy and sell stocks per some predefined rules. And if their shop is any good at running experiments, they can do this with as little friction as possible.

The same holds true for other industries: the ability to quickly and efficiently conduct experiments is the key to success in the ML/AI world. Having more data and using fancier tools only goes so far. Your researchers need to be able to test a lot of ideas – without burning through a budget of time and money – and see which ones work out.

All of this means you’ll need to develop policies and procedures, establish areas where people can test models, and make it easy to pull and understand training data. This will help your researchers focus on actually building models, which should lead to faster time-to-market and better results.

This leads us to data infrastructure, which I’ll cover next.

  1. The stock market is also the ultimate real-time feedback system (with with stand-up comedy running a close second), which means you know very quickly how well your strategy performs. ↩︎