An evolutionary artificial neural network time series forecasting system

Abstract

Artificial Neural Networks (ANNs) have the ability of learning and to adapt to new situations by recognizing patterns in previous data. Time Series (TS) (observations ordered in time) often present a high degree of noise which difficults forecasting. Using ANNs for Time Series Forecasting (TSF) may be appealing. However, the main problem with this approach is on the search for the best ANN architecture. Genetic Algorithms (GAs) are suited for problems of combinatorial nature, where other methods seem to fail. Therefore, an integration of ANNs and GAs for TSF, taking the advantages of both methods, may be appealing. ANNs will learn to forecast by back-propagation. Different ANNs architectures will give different forecasts, leading to competition. At the end of the evolutionary process the resulting ANN is expected to return the best possible forecast. It is asserted that the combined strategy exceeded conventional TSF methods on TS of high non-linear degree, particularly for long term forecasts.