Wednesday, December 8 2021
10:00am - 12:00pm
Masters Presentation
The importance of parameter specification in Bayesian ARMA Models

Autoregressive moving average models are widely used for forecasting due to their intuitive structure and ease of implementation. Bayesian ARMA models are gaining in popularity as they allow the forecaster to directly incorporate prior beliefs about the parameters and the series. Parameter estimation for ARMA models is carried out by the Box-Jenkins method, which uses autocorrelation and partial autocorrelation functions to identify the autocorrelation structure of the series. This method is far from certain, and the true value structure of the autocorrelation is often unknown. This study looks at how Bayesian ARMA models forecast differ when the autocorrelation structure in the models is known to be wrong. Several time series were simulated, estimated with Bayesian inference using an incorrect specification and compared the forecasts to correct models. The Results show that parameter specifications that are close produce similar forecasts although with wider credible intervals and that overfit models have more consistent forecasts with the correctly specified models than underfit ones.
Speaker:Heath Lancaster
Location:Zoom see email

Download as iCalendar