|
Abstract:
|
This talk will discuss a fundamental principle in time series analysis; well-fitting models produce assumption fitting residuals. We propose a method of testing Gaussianity and correlation concurrently for a given set of model residuals. In current automated fitting routines (and most time series classrooms) the status quo is to pass residuals through individual diagnostic checks. Residuals that pass all tests are deemed to be satisfactory and the final model choice relies on AIC/BIC, likelihood, parsimony, gut instinct or some other criteria. We take a different approach; our methods devise a single test that allows for concurrent testing. The main improvement here is the ability to rank sets of residuals by which are closest to the model assumptions. The framework used is a "black box" fitting methodology where residuals are passed through in an automated fashion and results are independent of the model.
|
