Using Genetic Algorithms to Forecast Financial Markets


Traditionally, prediction engines used to forecast financial markets have employed deep learning or feed forward neural networks to optimize weights and derive an optimal prediction. Neural networks have weighty activation functions and permutationally many parameterization functions losing adaptability in the process. In an effort to create a highly adaptable entity we employed more evolutionary based concepts: specifically genetic algorithms (GA). This approach reduces dimensionality of the feature space and enhances the generalizability of the classifier in parameterization. We represented each chromosome as a projective and reactive trading philosophy with a set of input parameters populated by technical analysis variables and a sentiment variable. The calculation of sentiment accounted for complex linguistic systems in unstructured data using relational third factor dependency trees. The system propagated vector error through multiple evolutionary strategies. Pattern propagation in new patterns (like the crash of 2008), convergence, and adaptability was tested using multiple data sets from a wide range of financial markets. The GA based system displayed a more targeted approach in the refactoring and parametrization of weights due to the maintenance of a single “activation” function maintaining vector errors beneath 1.3E-3 per frame. This, in combination with the inherent nature of GAs, created a highly adaptable projective entity best suited for low latency trading in highly volatile financial markets.