So I sat down with an old friend who was literally the equivalent of a rocket scientist at a major F20 materials company for over a decade or two - now retired to run a predictive trading hedge fund. We talked about tagging and how it could play out in his trading tool. I took careful notes, and, this is in essense what he was driving at in terms of his approach to a trading strategy.
His forecasting model, like most others, is designed to increase in accuracy with the magnitude of market swings. The ability to predict accurately both "fat tails" of the return distribution is critical for developing profitable trading systems. Many current forecasting methods use total forecasting accuracy over the entire return range of the distribution.
Simple trading functions (or those that will fail consistently, such as Always Up or Always Down models) will predict one side of the return distribution perfectly, but be completely wrong for the other side. These stiff functions can be used as the basis for building adaptive, regime switching tradinig systems that are used in several trend following methods. Clever switching strategies can result in significant cumulative returns, at the expense of high volatility due to the stiffness of the function. In contrast, his approach is to develop soft trading functions that are more balanced in their ability to predict tha fat tailson both sides of the return distribution. He uses proprietary methods based on information theory and genetic algorithms to discover these functions.
In order to discover underlying market structure, he constructs a large set of features that broadly span (financial factors, time, statistical metric) space. He thinks tagging financial information will help make this intial data capture step much more reliable and robust since errors caught in this first steps can lead to signals that could be 100% off.
No comments:
Post a Comment