Like Nate Silver, only less accurate
These trendlines are fitted using a linear state space model. The underlying state variables are the primary votes for each party in each State and Territory. The model fits a simple local level model to each component, but it uses a factor model for the covariance of each week's innovations. Estimates of national votes and TPPs are made using population weights and last election's state-by-state preference flows.
Each polling source has a fixed house effect in the form of a Gaussian error distribution. Elections are assumed to be perfect observations. I match the timing of each polling sample, so that e.g. Newspoll's quarterly averages are modelled as noisy observations of the underlying data averaged over a quarter.
The complete code and data for the model are available on GitHub.
There's another statistical polling model over at Mark The Ballot (now back online, woohoo!). The Poll Bludger and Dr Kevin Bonham both publish regularly updated polling aggregates which they put together using their own judgement.
I've implemented Google Analytics, because I'm curious to learn more about how people are using the site. So be aware that, by default, data on your use of the site may be collected in that way. You can opt out of this by blocking cookies, by using Google's own opt-out tool, by browsing the site in private mode, or by installing Ghostery. In the future, this site will also store a cookie to help it remember your preferred graph settings. Again, feel free to opt out.