Election results madness

With help from the AEC scraping code by Mick McCarthy here, I've tried to predict the election results once the "non-ordinary" votes (early, pre-poll, postal and provisional) have been counted. These tend to go more for the Liberal party, so seats that Labor is currently ahead in might be lost. 

I’ve assumed that:

• all ordinary votes are in (only non-ordinary votes left), 
• the proportion of formal votes remains the same as in 2013 for all electorates, 
• the number of non-ordinary votes all increase proportionally with each other, 
• and the changes in 2 candidate preferred votes between voting types (I call this "bias") in each electorate stay the same between 2013 and 2016 (e.g. if the Libs get a 2% bounce in postals in Denison in 2013 compared to their ordinary vote result, that same proportion holds in 2016).

It turns out that William Bowe of the Poll Bludger, my favourite election analysis page, pipped me to the post with his analysis here, but I thought I'd run mine anyway. And it gave similar results, but importantly different enough to get the Coalition over the line. Here are the number of votes that the LNP (Liberal/Nationals party, well, Coalition) are expected to be ahead or behind in the closest seats:

Capricornia LNP +547

Chisolm LNP +1230

Cowan LNP -580

Flynn LNP + 1469

Forde LNP +86

Gilmore LNP +713

Herbert LNP -315

Hindmarsh LNP -786

Melbourne Ports LNP -1224 (I didn't do 3CP analysis though...)

Petrie LNP +1447

The Poll Bludger's analysis has them behind in Forde by 18 votes - my extra seat gives them 76 seats, a majority in their own right, whereas his 75 seats is not enough.

I then had a play with adding random variation, relaxing the assumptions of the proportion of formal votes staying the same and the amount of bias staying the same. I assumed that the overall variance between electorates for these properties stayed the same, but that a little (10%), half (50%) or all (100%) of the variance was due to random variation and not the specific effect of being in one electorate or another. The less important electorates became, the less likely a Coalition victory - for a little, 66% of simulations resulted in a win; for half, 34% and for full random variation, only 20% of cases.

Edit: Using 2010 results instead give us similar results, though slightly worse for the Coalition - 76 seats for the Coalition without variation, and 53%, 20% and 16% in the three scenarios described respectively.

Edit #2: Using 2007 results is more difficult because of seat redistributions, but we can do it if we assume that Liberals and Nationals will experience the same swings in Capricornia and Flynn. The model also gives 76 seats for the Coalition without variation, and 91%, 49% and 27% for the scenarios.

The model also pointed to some seats that could potentially be very close (winning 2PP < 50.2%) in these situations, so watch for these potentially coming into play if things get even more interesting:

likely:

Batman (VIC)       

Longman (QLD)

possible:

Banks (NSW)

Dickson (QLD)

Dunkley (VIC)

Griffith (QLD)

La Trobe (VIC)

Lindsay (NSW)

Robertson (NSW)