Um modelo linear com amortecedor de tendência dinâmico para previsão Bayesiana de séries temporais

Abstract

We investigated the performance of a dynamic linear model where the trend damping parameter has a time course given by a normal probability distribution. The objective is to determine if the inclusion of a dynamic evolution of the damping parameter improves the predictive performance of the model compared to existing polynomial models, in particular additive trend model Damped Holt. To evaluate this new proposal, we developed a simulation study and application on data from international competition M3, analyzing the performance of the model. They were simulated order polynomial series with two different observational values ​​for variance and variance of the states of progress. The study results suggest that the proposed model can get a marginal predictive gain over existing polynomial models in much of the parameter space. With data from international competition M3, several series with different characteristics were analyzed. The predictive function K steps forward was evaluated after a period of adjustment of the model to the data. For the estimation of the parameters of existing models, we used the technique of multiprocess class I and to estimate the parameters of the new model was employed to minimize the measurement error SMAPE. The new model has all the evolution of the states in analytical way and any type of simulation is required for parameter estimation. The limitation for the new model emerges as the study parameter related to zero slope. In this case the model is considered inappropriate and further studies are needed to work around this problem.