Non-Stationary Mixed-Frequency Variables in Time Series Models: A New Design Approach

Abstract

Mixed-frequency variables may encounter problems of non-guaranteed steady-state in time-variant state-space system during temporal disaggregation, forecasting or nowcasting. The instability of state-space system directly affects the accuracy of prediction. This thesis aims to develop a new design framework to model non-stationary mixed-frequency variables in time series models. We introduce a periodic constraint to control the instability of time-variant state-space system for non-stationary mixed-frequency variables. Our proposed periodic constraints in time-variant state space system are originated from temporal-aggregated constraints themselves. We fully utilize the binding conditions of both unobserved and observed temporal-aggregated conditions to generate the bounded periodicity of Kalman gain, control the instability of time-variant state-space system and improve the accuracy of temporal prediction. Such constrained state-space system for mixed-frequency variables we proposed is implementable with a conventional Kalman filter.