Discount Bayesian models and forecasting
This thesis is concerned with Bayesian forecasting and sequential estimation. The concept of multiple discounting is introduced in order to achieve parametric and conceptual parsimony. In addition, this overcomes many of the drawbacks of the Normal Dynamic Linear Model (DLM) specification which uses a system variance matrix. These drawbacks involve ambiguity and invariance to the scale of independent variables. A class of Normal Discount Bayesian Models (NDBM) is introduced to overcome these difficulties. Facilities for parameter learning and multiprocess modelling are provided. Unlike the DLM's, many limiting results are easily obtained for NDBMM's. A general class of Normal Weighted Bayesian Models (NWBM) is introduced. This includes the class of DLM's as a special case. Other important subclasses of Extended and Modified NWBM's are also introduced. These are particularly useful in modelling discontinuities and for systems which operates according to the principle of Management by Exception. A number of illustrative applications are given.