Lookback Period, Epochs and Hidden States Effect on Time Series Prediction Using a LSTM based Neural Network

Abstract

Forecasting time series problem occurs in various subject areas. Recently neural network techniques have been used for solving such tasks. However, they have not been sufficiently studied. The article explores the influence of the lookback period, the training epochs, and hidden state dimensionality in forecasting time series using long short-term memory. Numerical experiments with example financial data show that using more lags does not improve the results. Such a study of model parameters is important for their proper selection.