A comparative study of singular spectrum analysis, neural network, ARIMA and exponential smoothing for monthly rainfall forecasting

This paper investigates the accuracy of several forecasting methods for monthly rainfall forecasting. First, we study the feasibility of using the Singular Spectrum Analysis (SSA) to perform rainfall forecasts. When the time series data has the outliers, SSA might results in misleading conclusions, and thus robust methodologies should be used. Therefore, we consider the use of two robust SSA algorithms for model fit and model forecasting. The results of these forecasting approaches are compared with other commonly used time series forecasting techniques including Neural Network Autoregression (NNAR), Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS) and TBATS. The performance of these conjunction methods is compared in terms of accuracy for model fit and model forecast, using the monthly rainfall data from four rain gauge stations in Guilan province of Iran as the case study.

Keywords:

Contaminated data , Forecasting , singular value decomposition , Time series