The signal and the noise – Nate Silver in conversation

For those few not in the know, Wonkfest closing speaker Nate Silver is a genuine wonk icon.

His book, The Signal and the Noise, details a journey through Bayesian statistics that has revolutionised baseball and thrown a curve ball at political commentary and electoral prediction. He famously predicted forty-nine of fifty states in the 2008 US election – and maintains detailed predictive models of sports competitions and elections on his site, Five Thirty-Eight.

His workrate is legendary – the opportunities to hear him speak in the UK are astonishingly rare. So, despite coming at the end of two packed days at Wonkfest, a full-to-bursting auditorium were eager to hear what he had to say on two of the key themes of the day – politics and data.

In conversation

Nate was clear that UK politics was “way more messed up” than in the US – the sheer number of gyrations and reversals over brexit seems from afar like overcompensating for something in the past. Some would say the US are also in a constitutional crisis (which is fair to argue) but people are motivated to vote and to change things – whereas in the UK voters seem tired and burned out. Until voters weigh in, in 2020, we can’t know for sure.

Mark Leach led him onto the delicate matter of impeachment – although the lower house (representatives) will pass it, the upper house (senate) is Republican controlled and is unlikely to vote for it. But there are only three or four other datapoints (depending on how you count “serious” impeachments) so it is difficult to predict.

Public testimony – due in November – will be a powerful stage. We need to watch where Republicans may start to wobble. Perhaps the dam will break. Roughly 50 per cent support impeachment, 44 per cent are opposed based on Nate’s reading of polling trends. The audience remembered brexit polling, and shuddered.

Will it end in tiers?

On the Democratic primaries – Nate sees different tiers in the primary, with Biden and Warren at the top, and both quite far to the left in US terms, with Biden (76 year old) more moderate. Warren, meanwhile, is a “wonky populist” with many broad left policies worked out in full detail.

A tier down sees Sanders (a different flavour of working class left populism) who is popular but divisive. Buttigieg is an interesting character, who is a contender particularly in Iowa (the first scheduled primary).

Hillary Clinton? “Tier voters are tired of her”.

Primaries have only been a thing in the US since 1972, prior to this a literal smoke-filled room saw presidential candidates selected by those with money and power. But with three of four main candidates (Warren, Saunders and to a great extent Biden) are disconnecting themselves from party elites.

But could any of these candidates beat Trump? Yes – the candidate is not as important this far out, and Trump is not very popular (with a not-catastrophic but still historically poor) 41 per cent rating. People in the US may not realise how controversial the president is internationally

A good poll

Five thirty-eight recently issued their pollster rating – in essence more transparent and better documented polls perform better, and telephone polls perform better than online polls in the US. In polling you get what you pay for.

Voting in the US has become very deterministic by race and location – 92 per cent of Black voters selected Clinton last time round. Democrats are in cities, Republicans in the country. Polling has become increasingly accurate since 2016 – and, honestly, polling was within the margin of error in 2016, as it was for the brexit referendum.

Mark Leach asked why polling in the UK seemed so broken. Nate recounted his prior attempts to build models – he found that pollsters in the UK tend to “fight the last war”, for example adjusting the conservative vote up in 2017.

Multiple parties bring tactical voting which makes things very difficult, two party systems are much easier to predict. And UK campaigns are very compressed.

“Most Americans regard British politics as comic relief” – in the US people like Bercow and Johnson are figures of fun – though the endless flowcharts on brexit outcomes are a cult success.

Good use of data?

Nate Silver is worried about some of the ways “big data” is being used – he values his privacy highly and worries that he has no control over what data is used. There has been a fairly belated backlash over data collection by Silicon Valley giants. Any time data is collected there are trade-offs.

In education, there are always good ways and bad ways to use data (“a cliche”) but there is never just one question to answer. We can never know which is the best university for the average students. Institutions are often afraid of data because it may catalyse change – people should not worry about change per se but should be leery of those who claim data holds all the answer.

Mark described, carefully, the plans for TEF. Nate had noted his alma matter LSE had been rated “Bronze”, which did not make him happy. Pulling three unlike things into one rating does not make sense – it is better to keep satisfaction, outcomes, and continuation as three separate and very complicated measures.

I don’t like mushing several unlike numbers together when they are telling different stories

Different students from different backgrounds have different priorities – an all in one system like TEF does not address this.

In the US graduate earnings are less important than graduate destinations – but preferences change and investment banking is not what it once was. Salary does not correlate with other desirable outcomes from learning, and some fields are always more or less lucrative.

Questions

Attendees with links to the US were keen to tap Nate’s encyclopaedic knowledge of senate races and thinking on the unlikely popularity (with Republicans, at least) of Trump and whether Pence would be even more popular, but several questions tackled UK issues. For example Nate feels like as things are “fairly screwed up” in UK politics it may be time to consider electoral reform. He was winningly dismissive of his qualification to comment – though clearly not a fan of the Fixed Term Parliament act.

On opinion polls, Nate sees their purpose as to prevent journalists from entirely constructing their own narrative, a check against groupthink. Maybe, in the UK, there are limits to the usefulness of polls – though in most places they remain more accurately than opinion journalism.

In terms of data-driven decision-making advice – he cautioned that organisations need people who work with data and people who can enact change, but it was important to empower people to critique data driven decision. People using data in ways that confirm prior assumptions is a “sophomoric stage”.

And on data quality – in the long term, good descriptive data should produce good predictive outcomes. Five thirty-eight doesn’t use machine learning, their methods need domain knowledge (“this is why our UK election model sucked”). When we notice that a piece of data is potentially wrong we need to know when findings are illogical. He worries about models that predict well but you don’t understand why – these are often just “lucky variables”.

A final question asked how good quantitative skills are produced – Nate noted that classically you don’t get much statistics or data science in US schools. But universities are stepping in to address these needs – the calibre of students hired at five thirty-eight has gone “right up”. But teaching can focus too much on formal techniques, and not on “wrangling” techniques like how to ask a good question of data.