|Friday, February 16|
|CS06 Bayesian Applications||
Fri, Feb 16, 11:00 AM - 12:30 PM
Forecasting Periodic Accumulating Processes with Semiparametric Distributional Regression Models and Bayesian Updates (303524)*Harlan D. Harris, WayUp
Keywords: empirical bayes, bayesian modeling, distributional regression, generalized additive models, industry, application
Some business forecasting problems take the form of a count that accumulates up to a deadline, such as total monthly sales of a product, or student signups to an educational event before it starts. In this talk, I'll describe two real-world examples of this problem, and then will describe a Bayesian approach that combines a prior distribution with a principled, statistically honest extrapolation method. A key piece of this approach is a class of semiparametric distributional regression models called GAMLSS -- Generalized Additive Models for Location, Scale, and Shape. These models can flexibly describe the shapes of the empirical priors and temporal accumulation functions, and allow posterior predictive intervals that become more precise over time as information accumulates. This is important for nontechnical users of business predictions -- the inevitable errors in overly precise predictions lead quickly to mistrust.