Forecasting Time Series with Artificial Neural Networks

Abstract

Artificial Neural Networks (ANN) are used as universal approximators and are getting widely adopted in several working fields such as finance. One of the problems that can be addressed using ANN is the forecasting of time series. Time series forecasting is known to be a difficult problem, often requiring expert knowledge, and can be applied to problems including predicting stock value, sales forecasting, and inventory. We explore how the sparsity that occurs in trained ANNs can be used to generalize the network topology to any Directed Acyclic Graph (DAG), instead of running on a layer based architecture. Allowing the ANN to run on any DAG allows it to use the full capabilities of the input, and intermediate values. We show how both the Feed-forward Neural Network (FNN) and Recurrent Neural Network (RNN) topologies can be generalized to this variant, in both training and prediction. Finally we train these network architectures on benchmark problems and use them to forecast time series, where we demonstrate a powerful algorithm to predict the stock market values one day ahead.