We'll look at monthly Google search query hit counts in the UK and NZ for the term "queen's birthday".
After a rapid EDA, supplemented by a touch of extrogenous research, we'll fit an Normal Dynamic Linear Model (NDLM) to tackle three common timeseries tasks in a Bayesian manner:
Find the distribution of the number of hits at time t given
- the available information up to and including time t (filtering)
- all available information (up to and beyond) time t (smoothing)
- only the information available up to time s for some s < t (prediction)
The fitted model describes the data remarkably well and accurately predicts 12 months into the future! In spite of this, all models have caveats and failure modes: we close the project by highlighting assumptions and future events that might cause the modelling to diverge significantly from reality.
Two files are of particular interest:
projects/gtrends_project.Rmd: The R Markdown version of the project.projects/gtrends_project.pdf: A pdf version of the output!
A summary plot of the model's filtering and (12-month) prediction distributions:
