Measuring and Combining Edges (Part 3): Ensemble Learning
If markets are non-stationary and edges appear and disappear then we need something more complex than simple prediction models used in other fields. If physicists cannot definitively place the exact location of an electron, than what hope do we have of doing a one-time regression model on the S&P500 and basing our investing decisions on that? Yet this is what a lot of people like to do—borne out of our primitive abilities to understand randomness and complex systems. A good model to address the nature of financial markets must be: 1) dynamic 2) adaptive 3) complex (to account for conditional situations) and 4) recursive. There is only one method that can accomplish these goals effectively and that is an ensemble learning method.
Ensemble learning methods are analogous to the way decisions are made within a large hedge fund. In this case a research team interacts with portfolio managers who in turn interact with traders who themselves in turn answer to risk managers. The process is also circular where the risk managers talk to the portfolio managers, who in turn alter decision making. While the quality of the decisions made in this process can vary widely depending on the firm, the process itself is quite sound and would be highly useful if the decision inputs were accurate. This method makes tracking thousands of stocks and dozens of major markets a much easier task and likely more successful than having one portfolio manager do all the work themselves. The accuracy and quality of the decision-making are likely to be superior. Furthermore the depth and complexity of analysis are much less constrained within this large scale interactive format. No portfolio manager operating alone could account for all of the myriad factors that affect even a single stock in real-time. Even if time was not an issue, our own mental capacities are incapable of handling too many competing variables at once.
This analogy holds true for mathematical prediction problems- no single regression will be able to capture the factors that affect the S&P500 in one equation without serious compromise. Even non-linear methods and neural networks are incapable of handling all of the mitigating factors at once. This is why many higher order prediction models often fail out of sample– they learn everything there is to know about how a batch of variables interact without knowing much about how each variable actually works. They also fail to generalize. In contrast a commitee of models let by a single managing model (ie the portfolio manager) is far more likely to succeed in this task. Such a committee is exactly what the field of ensemble learning is all about and represents the true key to good prediction.