A dynamic factor model for multivariate water quality time series with trends

Abstract

A general dynamic factor model for multivariate nonstationary processes which simultaneously takes into account the time-lagged correlations and the linear trends of time series is developed. The distinct advantages of this model are the reduction of high dimensionality of original variables, the facilitation of direct estimation of the linear trends, and its capability for forecasting. Application of the model to water quality data for six constituents collected at a station on the Arkansas River shows that the estimated values of linear trends are close to those estimated by using the nonparametric seasonal Kendall method. Both model fitting error and forecast error variances are relatively small.