|
Abstract:
|
We are considering the problem of predicting social unrest using multiple sources of information as model features. These include social media, blogs, news, and currency exchange rates. The ground truth for fitting these models is manually created, and we are typically in a situation where there are many fewer observations than possible predictors. In addition, we hypothesize that over time, different predictors may be more important. We propose the use of dynamic models to forecast these complex non-stationary time series. The time-varying parameters of these models allow for flexibility in short-term forecasting because the time-varying structure allows for parameters to evolve with changes in a system over time. The dynamic linear model (DLM) assumes random Gaussian error and is a well-studied problem in terms of both model fitting and effect selection. Conversely, the dynamic generalized linear model (DGLM) assumes non-Gaussian distributed data and is a challenging model to fit. We present a variable selection method for the dynamic logistic regression model to find a parsimonious representation of a high-dimensional model to predict the probability of a binary outcome.
|
