Machine learning and artificial intelligence are used interchangeably but they are essentially two different but inter-related approaches. The distinction I make when viewing the two is that artificial intelligence refers to the use of existing rules/formulae and memory in order to make decisions, while in machine learning we provide the system with inputs and outputs, and the algorithm figures out the rule or formula.
In investing, both approaches hold value, and based on news reports, it seems that almost every hedge fund or asset manager is trying to apply these techniques to improve investing — with good reason.
This recent surge in applying machine learning and AI techniques to investing is likely driven by the recent breakdown of a massive wall. While machine learning and AI algorithms have been around for decades (I recall my first internship 20 years ago where I worked on genetic algorithms for a software firm), the high cost of computing, and the complex programming required put these techniques out of the reach of the finance quants, who were more spreadsheet oriented. However, the availability of open source libraries such as Google’s tensorflow, which put incredible amount of computing power and algorithms into the hands of non-computer engineers has opened up a new way of evaluating and understanding investing.
Here is an illustration of how easy and effective machine learning can be in predicting asset price movements. After reading and borrowing heavily from online tutorials such as those available at Machine Learning Mastery, I used python and tensor flow to develop a basic machine learning algorithm to predict the movement of the VIX Index.
This example took me about an hour to create and run on my laptop. For this illustration I downloaded daily VIX data from 2004 onwards from the CBOE, separated the data into a training set and a test set, and ran a basic machine learning algorithm to check how good the prediction might be.
Here are the results:
The model was trained using the training data, and then tested using the test data. As you can see from the chart below, it appears that the algorithm found a decent relationship in the training data and was able to successfully apply the model to the test data as well.
Now lets look at a scatter plot the compares the actual value to the predicted value. Looks like the predict worked pretty well. With r-squared of 86% it appears to be a pretty good relationship.
This is a proof of concept and should not be used to develop trading strategies without further due diligence and understanding of market dynamics. However, I am excited about how such approaches can expand data driven investing by putting powerful computational tools into the hands of individuals with capital markets expertise.
If you are interested in discussing more, drop me a note at amit.sinha@focus262.com.
Also published on Medium.