Saturday 2 October 2021

Bear With Me

It's been a busy time for voting and elections generally, but it would be remiss of me to not discuss the big news this week.

That's right! It's Fat Bear Week!


Fat Bear Week is an annual voting contest run every year since 2014 (though in 2014 it was called Fat Bear Tuesday and lasted only one day). Votes are placed through an elimination bracket system to determine a winner, but the rules are vague as to what the voting is on.

Photos of the Bears of Katmai National Park are shown in before-and-after style of their pre-winter torpor weight gain, suggesting the votes should reflect the fattest bear, the most weight gain, the most proportional weight gain or something similar to this. Each of these could be more objectively measured than through public collective guesswork so I have to expect that people are really just voting for their favorite bears.

Whatever the intent is, the votes are already underway and wherever there's a poll there's an opportunity for a statistically dubious prediction based on publicly available (i.e. incomplete) data!

It seems the contest has been run over multiple platforms, originating on Facebook and currently held on the Explore.org webpage. This made past results a little difficult to track down, but I am running off the following as my data sources: 2014, 2015, 2016, 2017, 2018, 2019, 2020.

As a first port of call, it's worth considering the past winners to see if there are any patterns, and early on there are:
2014: Bear 480 ('Otis')
2015: Bear 409 ('Beadnose')
2016: Bear 480 ('Otis')
2017: Bear 480 ('Otis')
2018: Bear 409 ('Beadnose')
2019: Bear 435 ('Holly')
2020: Bear 747

Evidently there is a good level of consistency among winners with two bears taking out the first 5 wins. This is promising as it suggests there is merit in considering past form, but as the two most recent winners break this streak it is possible that both Otis and Beadnose are past their prime.

Holly, as it happens, was eliminated in the first round and Beadnose is not in the running in 2021. This leaves Otis and Bear 747 as favourites at this stage and, at time of writing, they are paired off against each other for the Round 2 vote. In a matter of hours one of these will emerge victorious and be the hot favourite heading into the finals. Otis is coming in on the back of a Round 1 victory against Bear 402 who debuted in the first competition in 2014, while Bear 747 is coming in fresh from a bye. And this leads to an interesting point about the history of byes in Fat Bear Week.

An interesting point about the history of byes in Fat Bear Week

Since its inception, Fat Bear Week has run four Round 1 match ups followed by the winner taking on a bye. There is no shortage of bears in Katmai to flesh out the contest to 16 contestants and remove the byes. The bye system does work well for a week-long contest but I have to believe that is a coincidence as the 2014 contest introduced this format despite lasting only one day.


My theory is that the bye system is for the drama, introducing new bears into the contest after the first couple of days are done. Interestingly, the byes do very well in later rounds. Two thirds of byes (19/28, 2021 results unclear) win their debut match up, and with the exception of 2018 every Fat Bear Week winner had a bye. (In 2018 Bear 409 'Beadnose' won, and although not a bye that year has been one twice before this).

There is also a very select pool of bye bears. Over eight years only ten bears have held a bye:

Bear 32 ('Chunk') - 2017, 2019, 2021
Bear 128 ('Grazer') - 2020
Bear 409 ('Beadnose') - 2014, 2015 (won)
Bear 410 ('Four-Ton') - 2014, 2015, 2016, 2017
Bear 435 ('Holly') - 2016, 2018, 2019 (won), 2020
Bear 480 ('Otis') - 2014 (won), 2015, 2016 (won), 2017 (won), 2018, 2020
Bear 503 ('Cubadult') - 2019, 2021
Bear 747 - 2014, 2015, 2016, 2017, 2018, 2019, 2020 (won), 2021
Bear 856 - 2018
Cub of Bear 132 - 2021

This may provide a hint as to the success of the bye bears. Perhaps, to maximise the drama of Days 3 and 4, public favourites are preferred to fill the byes.

This may give us some insight on who to watch in 2021. Yesterday Cub of Bear 132 won from a bye, but Bear 151 'Walker' beat the other bye, Bear 32 'Chunk'. Currently the remaining byes (Bears 503 and 747) are lagging behind their rivals, which would make this only the second time that less than half the byes progressed. The other case was in 2018, when former bye Bear 409 'Beadnose' ultimately won.

While we're looking at form...  

There are a good number of bears who have competed across multiple years, but due to the elimination bracket system that's no guarantee they've run against each other. The table below compares all multi-contest competitors (2014-2020) and identifies those that have met more than once.


If the results of past contests were perfectly predictive of later ones, we would see this in the golden examples above--Bears 32 and 747 have been run against each other five times and we would expect the same bear to win each time. If the past has no bearing, we would expect about 50% of the rematches to give a different result. This latter scenario is closer to the truth:

  Bear 32    5 matches with Bear 747    2 wins, 3 losses
  Bear 32    2 matches with Bear 856    2 wins
Bear 151    3 matches with Bear 480    1 win, 2 losses
Bear 151    2 matches with Bear 856    2 wins
Bear 402    2 matches with Bear 854    2 wins
Bear 409    2 matches with Bear 410    2 losses
Bear 409    3 matches with Bear 480    2 wins, 1 loss
Bear 410    3 matches with Bear 480    3 losses
Bear 435    2 matches with Bear 747    1 win, 1 loss
Bear 435    2 matches with Bear 854    1 win, 1 loss

Around half of the repeat matches have consistent results, and only one example (410 v 480) exists where three or more consistent results were recorded. Therefore while form seems useful in picking a winner it is not as reliable in resolving individual match ups.

Current state of play

At time of writing the brackets for 2021 are as follows.


Previous winners still in the running are Bear 480 ('Otis') and Bear 747. In a matter of hours one of these will be outpolled and the other will be the smart money for Fat Bear Champion 2021. This pair have only matched up once before, where they vied in the final for 2017 Champion, which Otis ultimately won. As just discussed, however, this isn't much of an indicator that Otis will win again.

Otis has the most wins of any bear (2014, 2016 and 2017), but 747 won in 2020, where Otis lost in the second round after a Round 1 bye. It is entirely possible that Otis's days at the top of the Fat Bear tree are behind him, having lost his first round every year since his last win--a streak he broke on September 30. This win was comprehensive (71,227 votes to 14,026), and there are no shortage of Otis fans on the web.

This makes the Otis-747 match up a tough one to call, but early results have Otis ahead at a rate of around 3:2.

Currently today's other race has 812 ahead of 503 by a ratio of 4:3. That will leave the semi-finals with Bear 151 ('Walker'), Cub of Bear 132, 480 ('Otis') and 812.

Semi Finals

Bear 151 ('Walker') has contested every Fat Bear Week since 2016, but with little success. He has never been a bye and only made it past round two once (2020). Walker is up against the newcomer 'Cub of Bear 132', who is too young to have either name or number yet.

Cub came in as a bye and may be a crowd pleaser with its youthful looks and baby fat, but only narrowly defeated Bear 128 ('Grazer') 40,968 votes to 32,251 (= 56.0%). Grazer was not a particularly strong contender based on previous Fat Bear Weeks or on a similarly close first round victory over 2019 winner Bear 435 ('Holly').

By comparison Walker made the semifinals last year and has had two decisive wins this year with 64,451 votes to 11,820 (= 84.5%) in Round 1 against Bear 634 ('Popeye') and 60,621 to 11,038 (= 84.6%) in Round 2 against three-time bye Bear 32 ('Chunk').

If this represents a strong support for Walker rather than a tough pool for Cub, it's hard to see how Walker doesn't make the finals.

Prediction: Walker beats Cub of Bear 132

Bear 812 is the expected winner of today's first poll. This bear suffered a first round loss in 2018 and second round loss to eventual champion Bear 747 in 2020. This year it pulled a decent win over Bear 131 (68,092 to 17,706 = 79.4%).

However, the expected winner of today's second poll is three-time Fat Bear Champion Otis who preformed even better in Round 1 against Bear 402 (71,227 to 14,026 = 83.7%). The form, performance and fame make this an easy win for Otis.

In the event Bear 747 beats Otis today, this is does not meaningfully improve 812's chances.

Prediction: Otis beats Bear 812

On their current polling numbers (83-85%) Walker and Otis may look well matched, but again Otis's history as a three times winner really sets him apart.

Prediction: Otis beats Walker, Wins Fat Bear Champion

This would make 2021 only the second time a non-bye bear has won, and like the first time it would be to a previous bye-holding winner. The second most likely outcome to my mind is an upset win by 747 today that puts him in Otis's place on the Fat Bear throne.


Friday 30 April 2021

Tasmania's Blind Election

The Tasmanian election (today by the time this is posted) is a bit of a flashback to an earlier age for me, both in terms of methodology and forcing me to log in to my long-disused facebook. Back in the early days I'd cobble together some unreliable statistics from dubious sources and see if they were any use in predicting an election. Since then the Infographinomicon has become a little repetitive in applying polling to the existing pendulum and looking at outliers.

This will not be possible for Tasmania in 2021.

During this election there has been a single poll, and this is not a particularly reliable poll for a number of reasons set out rather fulsomely by Dr Bonham here. An analysis of what the outcome would look like if this were accurate is also already provided beyond that link, leaving little for us to do over this side.

Other than these numbers, the only 2021 data is from February. Also, of course, Tasmania's legislative assembly has multi-member seats assigned by the Hare-Clark system which doesn't lend itself to a pendulum in the way most lower houses in Australia do.

Working in the Dark

Without good polling data, we have a golden opportunity to reach for alternative analogues for popularity and test these out. My first thought was to look at the number of hits on the Liberal and Labor homepages. Although this would mostly capture politically engaged voters and a good number of people comparing both parties' policies, it is possible that more hits indicate an engaging news section drawing in undecideds or an enthusiastic support base sharing election materials regularly.

Unfortunately, neither party's page was pulling in enough traffic to show up on any of the trackers I consulted.


Not to fear--the prominent social media icons at the top of the party pages gave me another idea. Is is possible to look at social media metrics, such as Twitter, to get a feel for which party was doing better. Obviously a comparison of raw numbers like follower counts would include a lot of uncontrolled variables: political bias of Twitter users generally, different degrees of incorporation of twitter in other forms of communication, frequency and content of posts, and echo chambers could drive follower counts far more than voting intention.

Instead, what was needed was some kind of controlled measure. A value that was adjusted for engagement patterns and participation. An equation. A ratio.

Now for the benefit of those not involved in twitter--I am one of you--it's worth explaining what it means to be ratioed. Particularly in political discourse online, getting a lot of likes is a demonstration of support but getting a lot of comments or replies is generally not. A high comment-to-like ratio is generally taken as a sign you've either (a) said something grossly offensive or (b) expressed a negative opinion about some form of media with a large devoted following.

Obviously the extremes of being ratioed, where comments outnumber likes several times over, is unlikely to be a frequent occurrence on state or territory political tweets. But could even relatively tame like-to-comment ratios provide an analogue for political popularity? As it turns out... I don't know.

https://twitter.com/TasLiberal Poll Comment Retweet Like
Today N/A 5 7 13
This week N/A 86 143 237
15-23 Feb 52 200 18 13
17-23 Nov 52 N/A N/A N/A
18-24 August 54 N/A N/A N/A
3-10 Mar 43 10 7 7










https://twitter.com/tasmanianlabor Poll Comment Retweet Like
Today N/A 34 262 761
This week N/A 48 326 943
15-23 Feb 27 5 30 100
17-23 Nov 25 19 29 64
18-24 August 24 2 0 4
3-10 Mar 34 N/A N/A N/A

The data isn't there to assess this for Tasmania. I couldn't find dates for May polling to compare with to test this hypothesis, so these are omitted. The Liberal account tweeted too infrequently to find multiple tweets that aligned with a polling period. Labor only started tweeting last year, meaning there is no data for comparison to March polling and low engagement generally prior to the campaign.

Where data was available, it was too erratic to interpret. In February, Labor had an incredibly positive 20:1 like-to-comment ratio while the Liberals were below 0.07. In November, when polling was very similar, Labor's ratio was a more reasonable 3.37. There was also a lot of variability in the posts--high enagement seemed to align with retweets of federal political material which would eclipse state-relevant tweets. Comments, particularly on new accounts like Labors, seemed to be positive from loyal supporters and don't indicate negative interactions the way they do on larger established accounts. And, as an American dictionary points out high numbers of comments can be positive even then though this is not aligned with the pejorative nature of the word "ratioed".

With larger social media interactions over a longer period of time, perhaps correlations might emerge. There could be an interesting thesis project for someone looking at twitter ratios as a reflection of public attitude that expands this to look at tweets compared with television ratings or brand sales. But for our immediate question of who will win today's election there is nothing to be gained here.

This is disappointing, as this could have led to analysis of individual politicians' twitters to determine their personal polling. But this had other problems too--the large number or retweets, the variable prevalence of personal tweeting and the number of candidates without an active twitter.

Facebook provided a tempting alternative, with more consistent messaging, extensive post histories and more descriptive engagement options such as anger or sadness (though it was not always clear if a response was angry at the poster or their opponent who was criticised in the post). However, this huge number of posts and un uncooperative interface made it too difficult to recover the necessary data.

Return to Simplicity

Instead, as time rapidly runs out, let us try something less inventive but more promising. Here is a table of every Tasmanian election of the 21st century:


2002 2006 2010 2014 2018 Minimum
Bass Lib Lib Lib Lib Lib Lib
Lib Lib Lib Lib Lib Lib
Grn Grn Grn Lib Lib  
ALP ALP ALP Grn ALP  
ALP ALP ALP ALP ALP ALP
Braddon Lib Lib Lib Lib Lib Lib
Lib Lib Lib Lib Lib Lib
ALP ALP Grn Lib Lib  
ALP ALP ALP Lib ALP  
ALP ALP ALP ALP ALP ALP
Denison/ Clark Lib Lib Lib Lib Lib Lib
Grn Grn Lib Lib Lib  
ALP ALP Grn Grn Grn Grn
ALP ALP ALP ALP ALP ALP
ALP ALP ALP ALP ALP ALP
Franklin Lib Lib Lib Lib Lib Lib
Grn Grn Lib Lib Lib  
ALP ALP Grn Lib Grn Grn
ALP ALP ALP Grn ALP  
ALP ALP ALP ALP ALP ALP
Lyons Lib Lib Lib Lib Lib Lib
Grn Grn Lib Lib Lib  
ALP ALP Grn Lib Lib  
ALP ALP ALP ALP ALP ALP
ALP ALP ALP ALP ALP ALP

In the 'Minimum' column it is assumed, perhaps unjustifiably given all of 2020 and the recent WA election, that the Tasmanian election will sit somewhere within previously established extremes. For example, in Bass and Braddon the Liberals have always won at least two seats, so this is assumed to be the case today also. Denison (now Clark) and Lyons likewise have always returned at least two ALP MLAs, and Denison/Clark and Franklin consistently elect a green (for Clark this seems almost assured unless there is a massive rebuke of the Greens as this would be the party leader).

Although on-the-ground reports suggest otherwise, the expected dominant theme of the early Tas election was COVID-19. Excluding the most recent and dubious poll, it seems after March 2020 when the pandemic was declared, the incumbent Liberals' popularity shot up over 50% on first preference--higher than the 2018 election. We might assume, then, that every previous Liberal seat is safe.


2002 2006 2010 2014 2018 COVID-19
Bass Lib Lib Lib Lib Lib Lib
Lib Lib Lib Lib Lib Lib
Grn Grn Grn Lib Lib Lib
ALP ALP ALP Grn ALP  
ALP ALP ALP ALP ALP ALP
Braddon Lib Lib Lib Lib Lib Lib
Lib Lib Lib Lib Lib Lib
ALP ALP Grn Lib Lib Lib
ALP ALP ALP Lib ALP  
ALP ALP ALP ALP ALP ALP
Denison/ Clark Lib Lib Lib Lib Lib Lib
Grn Grn Lib Lib Lib Lib
ALP ALP Grn Grn Grn Grn
ALP ALP ALP ALP ALP ALP
ALP ALP ALP ALP ALP ALP
Franklin Lib Lib Lib Lib Lib Lib
Grn Grn Lib Lib Lib Lib
ALP ALP Grn Lib Grn Grn
ALP ALP ALP Grn ALP  
ALP ALP ALP ALP ALP ALP
Lyons Lib Lib Lib Lib Lib Lib
Grn Grn Lib Lib Lib Lib
ALP ALP Grn Lib Lib Lib
ALP ALP ALP ALP ALP ALP
ALP ALP ALP ALP ALP ALP

This assumption alone gives the Liberal party its majority back, but only leaves three gaps. In Bass the Liberals are at their 21st century maximum, so the last seat is between a Green candidate and Labor's Jennifer Houston (it is taken for granted the one Labor seat guaranteed will be Michelle O'Byrne as deputy Labor leader). The same polls that have Liberals above 2018 levels keep Labor under 30% first preference, which would be likely to favour the Greens who only failed to win a seat once.

The other two gaps are tougher to call, but the fundamental question is whether we are really seeing a return to 2014 levels of Labor losses where they polled 27.33%. The reliable but dated polling suggests so, but it's anyone's guess. I'm going to live dangerously and suggest this is the case.

One other point worth considering--in Clark the high profile, now independent Sue Hickey may win a seat. As a former Liberal I would suggest she might replace a blue seat, but others point out her campaign is quite attractive to the left and could replace a red or green. For simplicity I'm sticking to my above guesses for the following final prediction:


Prediction
Bass Lib
Lib
Lib
Grn
ALP
Braddon Lib
Lib
Lib
Lib
ALP
Denison/ Clark Lib
Lib
Grn
ALP
ALP
Franklin Lib
Lib
Lib
Grn
ALP
Lyons Lib
Lib
Lib
ALP
ALP