The term machine learning is overused - Have to get specific with what techniques are used

 

The term "machine learning" is being overused by many investment managers. It is being bantered about by some as a special technique or sauce that will magically liven returns for any portfolio. Sorry, there is no special magic, but there is a broad set of technique that can help analyze data. The magic is still connecting data with analysis in different ways to provide new insights.

Better analysis can lead to better returns. (The detailed mind map graphic on different machine learning techniques is from machinelearningmastery.com.) It does a good job of breaking down the wide variety of tools within machine learning and data science. If some portfolio manager says he uses machine learning, ask two simple questions: What techniques do they use? Why are these techniques better than a simpler analysis?

I wish there was greater use and explanation of these tools when in economic graduate school, but the focus of the times was different. The emphasis was on building models based on theory with the hope of testing hypotheses and making predictions. The new emphasis on a very simple level is on letting the data speak and then determining what it says that can be meaningful.

Data science today does not eschew theory, but places emphasis on prediction and extracting information. Theory is not thrown out, rather the focus is on extracting relationships from data suspected to be relevant. For the theorist, the key component for developing a machine learning solution is the feature engineering which determines what data are used. When discussing machine learning, focus on the feature engineering.

Using machine learning can be very attractive, but any expected results should be tempered by the fact that it does not automatically leads to excess returns.