I present experiments forecasting Canadian elections. There are several
novel aspects of this work not covered in previous studies. First, I
present
a methodology for creating a representative Twitter sample and validate
this sample against census data. Next, I describe a Vector Autoregression
with Exogenous Variables model that bases the forecast on existing
election
polls. A key difference from prior work is that I use the covariance error
to
monitor the forecast accuracy, detecing if the forecast is wrong before
the
election occurs. I test the model on the 2015 Canadian federal election,
not only accurately forecasting national and provincial results, but
outperforming
traditional polls. I'll then show how I use the person's location plus
census data
to derive a detailed demographic model of who is supporting a candidate.
Using
this model, I'll explore some key turning points during the 2015 election
and offer
some analysis as to what happened and why. Finally (time permitting) I'll
discuss
some strategies for moving away from election polls as the output
variable,
demonstrating a novel bootstrapping technique using American Idol data.