|
Abstract:
|
The class of stochastic ARMA models extends the classic ARMA time-series models by replacing (or smoothing) the deterministic relationship between target and regressors in these models with a conditional Gaussian distribution having a small controllable variance. As this variance approaches zero, a stochastic ARMA model approaches a classic ARMA model. We represent a stochastic ARMA model as a directed graphical model. In doing so, we benefit from the ability to apply standard graphical-model-inference algorithms during parameter estimation (including estimation in the presence of missing data), model selection, and prediction. The graphical model representation also offers a visual representation that is easy to interpret. We demonstrate how the graphical representation in this way lends itself as a conceptually easy way to handle cross-predicting time series, periodicity, and trends.
|
