Oz Politics Blog

Yesterday, Labor announced policies on or funding for cyber-safety, mentors for students, defence and IT projects in Northern Adelaide, congestion on Sydney rail, the botanical gardens in Canberra and a Wimmera-Mallee water pipeline.

The Coalition announced its forestry policy.

Newspoll: Today’s Australian had the latest Newspoll. The headline prediction was 54 to 46 for Labor.

The usual opinion poll graphs are here. You may need to hit the refresh or reload button on your browser to see the latest graphs.

Predictions: Simon Jackman predicted Labor would win 54 per cent of the two party preferred vote (with a 95 per cent confidence interval that the final result would be between 52 and 56 per cent. Drop that figure into Antony’s calculator, and it gives a prediction of 87 seats.

Simon’s prediction aligns with my perceptions of how the polls are trending. However, I discovered I am uncomfortable with the way both Antony’s calculator and mine translated two party preferred predictions into seat counts. We both used a simple, binary cut-off method. This can lead to very lumpy results. For example, with both of our online calculators 54.0 per cent for Labor yields 87 seats; but 54.1 per cent yields 90 seats.

A solution to my problem with the lumpiness of the online calculators came from a conversation I had with Antony Green a few weeks ago. Antony explained how the ABC does its election night seat prediction by summing the probability of the parties winning each seat (based on a booth by booth comparison function within seats). I decided to test a similar approach to converting TPP results to seat outcomes by summing probabilities. This approach allows a better recognition that individual seat swings are typically normally distributed around the national (or state based) uniform swing.

Rather than recording a seat as a win (1) or a loss (0), I have applied a cumulative probability function with a standard deviation of 2.3 percentage points. If a seat is sitting on the same margin as the predicted swing, it would have a 50 per cent chance of being won by both parties. If the seat was on a margin 4.6 percentage points below the predicted swing to Labor (ie. two standard deviations) it is 96 per cent likely to go to Labor. If a seat is sitting on a margin two standard deviations above the predicted swing to Labor, it has a 4 per cent chance of being won by Labor.

The results of my test of this concept follow. While Antony’s and my online calculators suggested 87 seats, the probability model suggested 88 seats. (There were also some minor differences between the probability model below, which is accurate to one decimal place, and the online calculators that round to two decimal places. In the binary cut-off model below, I have scored a seat as 0.5 if the swing equaled the margin).

But this leaves me with a dilemma. How do I compare election predictions from those commentators who have provided a prediction in terms of a TPP percentage, and those who have made a prediction on the basis of seat counts. My resolution is to covert the TPP predictions to a simple cut-off prediction from Antony’s calculator, and the probability prediction from my spreadsheet. The results follow.

As an aside, an advantage of a probability based model like the one above is that it allows for a direct comparison of seat win probabilities with the betting market. This allows for a better identification of under-priced and over-priced odds. In this context, having a view on the state based swings is also useful to model.

Next election, I wiill work on a probability based seat calculator, rather than the simple binary cut-off model I did for this election.

Update: Antony Green sent me the following email.

Just to let you know, my prototype calculator that was based on probability calculations produced 88 seats, not the 87 produced by the web version on the ABC site. That matches your table of calculations.

I get a +/- 5 seats on this prediction. The reason we didn’t do the probabilities on the web calculator was because of the extra calculation load imposed by the probability calculation, plus the fact that the +/- error was roughly the same whatever result came out of the calculator. And the only difference in Labor seats won tended to come out at 1 or 2 seats.

And Will From Kooyong sent me this.

I read your blog this morning, and it’s quite funny how yesterday I had decided to create a Monte Carlo simulator to predict seats that the likelihood of a win for the ALP. I used a std of of 3.05 (based on last election’s swings), and I tried your std of 2.3. The results I got were very close including the probability of seats changing hands.

Here is a post I put on Poll Bludger late last night when I ran it for a 54% TPP for the ALP, 3.05 for the std and 100,000 simulations. (I ran it for a 1m and it was very much the same).

http://www.pollbludger.com/?p=729&cp=3#comment-88637