What are machine learning hedge fund strategies?

There is a lot of resources being poured into machine learning strategies in the hedge fund industry. So what are machine learning based strategies?

Put simply AI/ neutral networks/ pattern recognition/ machine learning are all part of the same group of quantitative research processes. They are one part of the research process that leads to the actual trading strategy.

Most active investment strategies are based around a formal strategy, which aims to identify mispriced securities or to out trade other traders. The problem is how do you know you have the best strategy? I how do you know you have specified the strategy best?

Older quant strategies often started with a thesis, such as trading around directors dealings or earnings. The quant would research it on a semi quant/ semi discretionary basis, and if they believed they had an inefficiency that was exploitable, build a model to capture it.

The problem is, how do you know you have built the best model?

Simple, ask a computer to run as many combinations of the parameters that could describe the situation as possible and come up with clusters of similar situations/ trades and then build models around those optimal trades. This often leads to thousands of discrete models across multiple trading strategies. Take Directors dealing for example, some Directors may have a better track record than others, buying vs selling may offer different information, preceding earnings/ share price behaviour may have information in it, the job title of the director may have different information. Once you ask the computer to look down as far as the contextual track record of individual directors across thousands of stocks over 20 years of history, apart from the huge computing power need, you end up with thousands or tens of thousands of individual computer models.

These individual models aim to better capture the inefficiencies than one or a few models that are applied across all stocks and time frames. Ie you are going from a blunt one size sits all series of quant models to many discrete, tailored models. The ultimate objective being to capture the inefficiencies better than the next guy, whether he is a quant or not.