A0389
Title: Univariate time series forecasting using echo state networks: An empirical application
Authors: Alexander Haeusser - Justus-Liebrig-University Giessen (Germany) [presenting]
Abstract: The echo state network (ESN) is introduced for univariate time series forecasting. The echo state approach incorporates building a large dynamic reservoir, which models non-linear relationships between inputs and outputs to capture time series patterns like autocorrelation, trend, and seasonality. Even when the model is non-linear, training of the ESN is simplified to a linear model, estimated via Ridge Regression. A standard procedure for fast model estimation and selection is introduced. The modelling framework is suited for a wide range of applications, and an empirical analysis of real-world data from the M4 Forecasting Competition illustrates its modelling flexibility and forecast accuracy. The proposed echo state network can outperform or compete against state-of-the-art forecasting models from statistics and machine learning.
