Hacker News Reader

Comment by lazzlazzlazz | Hacker News Reader

It wasn't a blow out - if I recall, they gave a >30% chance for Trump to win. It seems perfectly ordinary for Trump to have won, then - "mild surprise" at most.

Their track record is still excellent.

View on HN

They also called out those who were claiming a trump win at < 2%.

They actually were called out for giving Trump such a high chance of victory by many. In particular, the Huffington Post accused Nate Silver and 538 of intentionally biasing their results to make it seem like Trump had a better shot.

Somehow in the months following the election the narrative completely turned around and now 538 were hacks for not giving Trump an even better chance.

Because of the winner takes all distributions in most states the result can flip very easily. If more states had proportional representation for the Electoral College the confidence level would be very high.

If they had just predicted the nationwide popular vote, Trump wasn't even close. The chance of him winning the popularity vote was probably in the order of < 2%.

Why do they get praise for their 2012 predictions but get a pass for their 2016 mispredictions? If they gave Trump a 30% chance to win and, as you say, it is perfectly ordinary for a 30% chance event to happen, then what is the point of their prediction? What is the point of listening to someone who can never be wrong?

Yes, the projections are essentially useless. Nassim Taleb does a great job of laying this out - basically, the projections create a false sense of certainty onto a highly uncertain set of events.

https://towardsdatascience.com/why-you-should-care-about-the...

https://math.stackexchange.com/questions/1911110/what-is-nas...

If I have the ability to predict coin tosses correct 55% of the time, that would be incredible - an almost unreal talent. And I'd be wrong 45% of the time.

For any one prediction, it's hard to say that a probability distribution is "wrong", but across many predictions, we can definitely evaluate the accuracy of a predictor and quantify whether they're useful or not.

538 has an excellent track record.

I know Mythbusters isn't the most scientific, but when they tested this, although with a small sample size so it could be off, if you choose Tails every time you should hit well over 55% based on the weight distribution of the coin.

They can be wrong. The problem is that you are looking at a single data point. Judging their performance requires viewing all of their predictions, across all elections, and seeing how often their predicted probabilities match up with outcomes. Do the people they give a 30% chance of winning win 30% of those races? Do the people they give a 60% chance of winning win 60% of those races?

That is how you judge their performance, and from that lens they do quite well.

> It wasn't a blow out - if I recall, they gave a >30% chance for Trump to win. It seems perfectly ordinary for Trump to have won, then - "mild surprise" at most.

Revisionist history.

I was in San Francisco the months leading up to the election, employees at startups were refreshing 538 pretty much every minute waiting for their latest data visualizations to confirm their pre-existing beliefs that "the democrat's got this"

To them, the literal worst case scenario was an electoral college TIE. Which of course would have also resulted in a Trump victory, but an electoral college BLOWOUT? And now you want to find satisfaction in defining the semantics of a blow out when it is obviously for emphasis and effect? seriously.

It was hopium since the primaries, 538 lost their credibility.

I was also in San Francisco and the delusion was incredible. I was posting almost weekly on Facebook that people were overstating Clinton's chances of winning, and that the pollsters giving it a <2% chance were way off the mark. I have the evidence to prove this, trivially.

> I was in San Francisco the months leading up to the election, employees at startups were refreshing 538 pretty much every minute waiting for their latest data visualizations to confirm their pre-existing beliefs that "the democrat's got this"

True, and consistent with what I've said.

> To them, the literal worst case scenario was an electoral college TIE.

I don't care what San Francisco startup employees expected.

> And now you want to find satisfaction in defining the semantics of a blow out when it is obviously for emphasis and effect?

Trump's victory wasn't exceptionally large in the history of the electoral college. You can verify this yourself.

> 538 lost their credibility.

They gave Trump reasonable odds to win. He did. I'm not sure what you're missing.

I'm unsure what you are talking about, but as the other person stated, 538 gave Trump a 30% chance to win. If your colleagues interpreted that as a surefire thing then that reflects on you all poorly, not 538.

Reply on news.ycombinator.com