Friday, December 25, 2015

2015 Ehrlich Awards For Wrong Predictions

Secular seasons' greetings and best wishes for 2016 to all.  Since this site started a few years ago, I've developed a strange habit of posting something every Christmas Day. As I may be too busy playing chess badly in early January to post all that much around then, I've taken the risk of going early with this year's prize for the unwise, the Ehrlich Awards for the wrongest predictions in a field of interest to this site made in or concerning the year 2015.  The Ehrlichs are named for Paul Ehrlich, the ecological don of doom whose failed resources bet with Julian Simon and poor excuses for losing it (and litany of other false "scenarios") have given heart to those who snort derisively at the claim the world is rooned ever since.  For previous instalments, and to see the groundrules, just click on the Ehrlich Awards tab at the bottom.

As usual we briefly glaze over hard-to-quantify bogus puffery from politicians of all varieties, spearheaded this year by Bill Shorten's commitment to the National Press Club "that Labor will be defined in 2015 by the power of our ideas." In fact, Labor was largely defined as a mirror of the government it opposed: popular by default while Tony Abbott clung to power then deeply unfashionable once he was given the boot.  A few Labor policies attracted public attention (emissions targets, voting age, smoking excise) but it was not always the right kind of attention, and some of these were pilfered from the Greens.  Labor has been releasing policies but none are especially original and so far hardly anyone has noticed.  Mostly they smack of a strategy tried without success against John Howard: when you have nothing just trot out the usual, cover it with a sudden interest in technology and science and say you've discovered something new.

A stalwart of these awards, Bob Ellis, doesn't make the podium for 2015.  Events wrongly predicted by Ellis in a mostly worthy 2015 included that Labor would win government in NSW, that the Greens' Arthur Chesterfield-Evans would win a 20% swing in the North Sydney by-election (it was actually 0.36%),  and that Tony Abbott would resign as Prime Minister in February.  However, Ellis ruled himself out of contention for the big one by predicting the Queensland election rather accurately (it was actually 51.2% and 44-42-3) which scarcely anyone else in public life managed to do!  (Psephologists generally, while not ruling out a Labor win, expected the LNP to probably scrape back in.  From the other side to Ellis though, check out Yale Stephens ringing the alarm-bell back in 2012.)


This year's bronze award goes to everyone involved in this by-election preference beatup by the SMH.  An argument that a Liberal Democrat "preference swap deal" (which was actually a unilateral decision by the LDP, not a "deal") would significantly hurt the Liberals in Canning applied Senate logic to a House of Representatives by-election with ridiculous consequences.  The Liberal Democrats - who have little traction in WA anyway - were never going to gain much benefit from a favourable ticket draw in the way that they do on oversize Senate ballot papers, and also micro-parties have little success in directing their Reps preferences.

Ultimately the LibDems finished dead last in Canning with 0.58% of the vote, and over 72% of this flowed to the Liberals anyway.  This meant that at most 0.16% of the vote (or one thirty-third of Andrew Hastie's winning margin) went Labor's way as a result of the LibDem's preference retaliation.  However since 20% of LibDem preferences flowed to Labor in 2013 anyway (and more than that in 2010) it's probable that less than a quarter of the LibDem voters swung to preferencing Labor, and of those who did, some may have done so independent of the preference decision (but for the same reason - the Liberal pursuit of the micro-party over name confusion).  The net impact of this supposed "blow" and "problem for the Liberals" would have been somewhere between zero and maybe (though I doubt it) thirty votes.


This year's silver award is the property of one John Black, a former Senator and sometime writer for The Australian.  Black used demographic data from the Queensland election to try to project the NSW election and came out with a projected result that looked a lot like Queensland.  (Title: "Newman case shows us why a Foley win is still very possible").   Nothing like that eventuated, and the main reasons for this are instructive as an example of how not to do predictive psephology.

The first problem here is that while demographics can explain a lot of variation within elections, they don't explain a lot of variation between them.  If they did then different elections would resemble each other much more than they do, since demographics don't change very much.  We might think of different elections as having their own base 2PPs and then predictable demographic factors (rich areas voting Liberal, for example) causing a fair part of regional variation in the 2PP.  Black effectively assumed that if NSW and Queensland did not vary much in demographics they should not vary much in the base 2PP.  Black did acknowledge this problem to a degree, but didn't pay it nearly enough heed.

The second problem is overfitting.  While some demographic variables, such as income, obviously do predict local variations in voting, Black used vast numbers of variables, some of which would have only accidentally correlated with voting patterns.  Models that do this are known as "overfitted" because they describe the previous data very well, but are worse at predicting future trends.

It didn't help that Black's understanding of what happened in the Queensland election to cause polls to be wrong, was itself wrong.  Only about half a point of the 2PP predictive error in Queensland resulted from actual error in the primary vote forecasts; most of it was down to radical changes in preference behaviour. By the time NSW rolled round, many pollsters had wised up to this issue and dealt with it.  Neither election saw major changes in voting intention through the campaign.  Most elections don't.

Tom Westland reviewed the damage at Crikey.  (Unfortunately other excellent Westland pieces about Black's unsound prediction methods are currently unavailable as Westland's blog is set to private.)  Black's model not only overestimated the swing but his model did quite a bit worse than the hackneyed old uniform-swing model despite (in fact partly because of) its far greater complexity, even after accounting for it getting the overall swing wrong.


Worthy as the above are, the supreme gong for this year goes to shockjock Alan Jones, who in June last year told then Communications Minister Malcolm Turnbull in a fiery interview:

"There is no challenge to his leadership. They are suggesting Malcolm precisely because you have no hope ever of being the leader. You’ve got to get that into your head. No hope ever. But because of that you’re happy to throw a few bombs around that might blow up Abbott a bit. That’s what they’re saying."  (my bold)

There was, at that time, some overheated speculation that Abbott was going to be dumped after less than a year in office and I pointed out some reasons why it didn't really add up.  Jones was right to also be sceptical of the immediate push, but forever is a very long time.  The obstacles to returning Turnbull to the leadership disappeared over the next fifteen months, while the obstacles to retaining Tony Abbott mounted.

Jones was hardly alone - Gerard Henderson was writing off the new PM in even 2011, and Google searching for "Turnbull will never be" (or similar) will bring up plenty more.  He was, however, the most prominent person to completely dismiss the now-Prime Minister's chances, and he did so although by that stage the evidence that Abbott was a dud was mounting.  This makes Alan Jones the winner of the 2015 Ehrlich Award.  (Speech! Speech!)

Feel free to add anyone I missed in comments.  I should be back around New Year's Eve with an annual site review, and I have some other slow-season stuff planned for the second half of January.

PS (26 Dec):  Somewhat relevant to the Black section above, I should also comment on this AFR article on social-media based predictions.  The claim made is that the social-media-based methods used predicted the Queensland 2015 and federal 2013 elections with 96% and 95% accuracy.  But I can find no evidence that the predictions were published or even derived in advance, and if you merely test a method on old data, that's not a prediction.  What you've done might actually explain the old result, or it might just be overfitted to it a la Black.  We do know that advance predictions using such a method failed dismally in the UK election, although that may have been partly because they also used polls as an input.  I suspect it is very difficult to eradicate the left-wing bias of social media in this sort of analysis.

1 comment: