AI Investing
How we use neural networks to predict stock returns
I am one of the most skeptical and conservative people out there. For example, I used my Siemens brick phone in 2012 when people already used iPhones everywhere. And when I hear people over-excited about AI these days, I stay a bit skeptical as it is natural for me. On the other hand, about 5 years ago I read a very unusual and rare book called “AI Investor” by Damon Lee. The book guided a reader step-by-step building an automated trading system using a simple neural network. From that moment I wanted a similar system of my own.
To be honest, the system didn’t do great even in his book, so the author was not too excited about the results. We all know the story of Hoover vacuum machines and his founder who only built a good vacuum machine after trying dozens of prototypes. I feel the same might be true about using AI systems for trading. You really need to keep building them until you arrive at something working decently.
I did my first iteration of the AI Investing system called Profit Prophet about a year ago and the system so far underperformed SP500. This is my first iteration and I didn't expect much. The network was trained to predict 1-year stock returns from the current point in time. The system is a 3 layered feed-forward neural network, trained on 10 years of historical data. The system uses 50 metrics per company. The examples of metrics are PE, PS, Debt-to-Capital ratio, Dividend yield, Beta, Margin etc. I also combined this network with similar networks to get an average and certain level of variance and stability.
Here is how the system looks like:
When the parameters are fed into the network they are normalized to be between -1 and 1. The network is then trained to predict one year return from various points in time during the last 10 years minus 1 year, and the network error is then computed as the network's prediction vs what the actual return was within a year from that point in time.
As I am writing this article, I am happy to announce that I trained a new network with certain changes from the first network design. I will know in about a year how well it performs! The new network is now available for free in the Profit Prophet section on Tickernomics website.