Meta-learning with Empirical Mode Decomposition for Noise Elimination in Time Series Forecasting

In time series forecasting, noise can have a cumulative effect on the prediction of future values thus impacting the accuracy of the model. A common method of machine learning in time series problems is to provide a number of past output values in the series so it can learn to predict the next value, however, other modes of time series forecasting also include one or more input series. This enables the application of the proposed technique in this study to provide additional meta-information to the model to guide learning and improve the prediction performance of the model. We identified the components of two time series datasets using empirical mode decomposition and trained a non-linear autoregressive exogenous model to compare its performance with the traditional approach. Two methods for processing the signal components for noise reduction were proposed and the result from the summed combination significantly outperforms the traditional technique.