high performance computing on graphics processing units: hgpu.org

Financial Time Series prediction attempts to model the behavior of financial markets using, among other things, tools like technical, intermarket, and fundamental indicators. Accurate prediction, however, is difficult for a number of reasons: financial markets are influenced, often in a non-linear, sometimes time-lagged fashion, by factors including interest and exchange rates, the rate of economic growth, and a number of industrial commodities. Neural Networks (NN) are a well-established method to attempt to conquer these difficulties [1]. Using unsupervised learning like backpropagation or the Levenberg-Marquardt algorithm NNs can be trained to model a market using historic market data; complex models, however, require the use of large NNs, the training of which requires large amounts of historic data – leading to long training periods. As the size of the network needs to be maintainable there can only be a limited number of inputs, leaving the network designer with the question what inputs to select from a large amount of technical indicators. The network topology also has a great impact on network’s modeling ability. There exist no exact methods for finding the "right" inputs or the "right" network topology: practitioners have to use heuristics until they arrive at a combination of inputs and topology that satisfies their requirements. Evolutionary Computing can provide a solution to this dilemma by maintaining a population of NNs with different inputs and topologies. However, as we’ve already mentioned, the time required to train a single network can be substantial, so the notion of training a whole population over a number of generations can render the algorithm infeasible for ordinary processors in terms of execution time. To overcome this problem we present an implementation which uses the GPU to speed up the search. The algorithm implements a new, NN-oriented, evolutionary search which is based on ideas borrowed from Grammatical Evolution (GE) [2], as well as other DNA-inspired concepts.