Stochastic State-Space Time-Varying Random Walk models have been developed, allowing the existing Stochastic State Space models to operate directly on irregularly sampled time-series. These TVRW models have been successfully applied to two different classes of models benefiting each class in different ways.
The first class of models - State Dependent Parameter (SDP) models and used to investigate the dominant dynamic modes of nonlinear dynamic systems and the non-linearities in these models affected by arbitrary State Variables. In SDP locally linearised models it is assumed that the parameters that describe system’s behaviour changes are dependent upon some aspect of the system (it’s ‘state’). Each parameter can be dependent on one or more states. To estimate the parameters that are changing at a rate related to that of it’s states, the estimation procedure is conducted in the state-space along the potentially multivariate trajectory of the states which drive the parameters. The introduction of the newly developed TVRW models significantly improves parameter estimation, particularly in data rich neighbourhoods of the state-space when the parameter is dependent on more than one state, and the ends of the data-series when the parameter is dependent on one state with few data points.
The second class of models are known as Dynamic Harmonic Regression (DHR) models and are used to identify the dominant cycles and trends of time-series. DHR models the assumption is that a signal (such as a time-series) can be broken down into four (unobserved) components occupying different parts of the spectrum: trend, seasonal cycle, other cycles, and a high frequency irregular component. DHR is confined to uniformly sampled time-series. The introduction of the TVRW models allows DHR to operate on irregularly sampled time-series, with the added benefit of forecasting origin no longer being confined to starting at the end of the time-series but can now begin at any point in the future. Additionally, the forecasting sampling rate is no longer limited to the sampling rate of the time-series.
Importantly, both classes of model were designed to follow the Data-Based Mechanistic (DBM) approach to modelling environmental systems, where the model structure and parameters are to be determined by the data (Data-Based) and then the subsequent models are to be validated based on their physical interpretation (Mechanistic). The aim is to remove the researcher’s preconceptions from model development in order to eliminate any bias, and then use the researcher’s knowledge to validate the models presented to them. Both classes of model lacked model structure identification procedures and so model structure was determined by the researcher, against the DBM approach. Two different model structure identification procedures, one for SDP and the other for DHR, were developed to bring both classes of models back within the DBM framework.
These developments have been presented and tested here on both simulated data and real environmental data, demonstrating their importance, benefits and role in environmental modelling and exploratory data analysis.