Parallel Implementation of a Forecasting Algorithm

Abstract

This thesis presents the implementation of a time series forecasting method in a High-Performance Computing (HPC) environment. This time series forecasting method integrates concordances and genetic programming (GP). The innovation in this work is the parallelization of a serial program. This is implemented on the OSU Cowboy High-Performance Computing Cluster (HPCC) using Message Passing Interface (MPI) based MPJ Express library. The two components which are parallelized in this thesis are concordance computation of a time series and developing an equation to predict the values of a time series. This parallelization was achieved by distributing the workload of the above-mentioned two components among the available processors for the algorithm to use on HPCC. For being able to parallelize and execute this algorithm on an HPCC, all the steps in the algorithm was executed without any manual intervention unlike in the originally proposed method. To make this algorithm run as a batch the genetic programming part of the algorithm was introduced with a stopping condition which also decided the accuracy of the prediction. So the implementation in this thesis allows the user to choose the accuracy of the predictor function by specifying the mean square error (MSE) limit and using this limit the predictor function development using GP can be stopped. The time taken by the model implemented in this thesis is inversely proportional to the MSE specified by the user.