An Analysis of Genetic Algorithms for Time Series Pattern Discovery

Professor Vasant Dhar

ABSTRACT

Genetic Algorithms offer a powerful mechanism for learning from data, based on ideas from Darwinian Natural Selection. In this talk, I will describe a genetic algorithm called GLOWER and its properties, and some results from using it to learn patterns from financial time series data. In particular, I shall describe the representation it uses for learning patterns, and results relative to those obtained via several other machine learning methods on a specific dataset. I shall also describe real-world experiences from using learned patterns to manage a global financial futures fund completely systematically. The latter are research in progress, including conjectures on when and why learning algorithms tend to “overfit” time series data.