Best Forecast One Week Before Election Day

[ursulahx]

I think my main gripe with the 538 model is how over-sensitive it is. It bounces around way too much in response to new data, so it ends up feeding the media hype of "she's up! He's down!" Surely the point of a prediction is it must anticipate this to an extent? A prediction that is constantly changing isn't a prediction any more, it's just another tracker.

(I know Polls-Plus is supposed to anticipate movement, but it's been almost as erratic as Polls-Only.)

[Terry the Fat Shark]

Just a note that 538 has it at 66, not 76

[Swedish Rainbow Capitalist Cheese]

Just a note that 538 has it at 66, not 76

He's talking about their forecast as it was on Thuesday though and not how it is at the moment.

[Terry the Fat Shark]

Just a note that 538 has it at 66, not 76

He's talking about their forecast as it was on Thuesday though and not how it is at the moment.

wow it's fallen 10% in 3 days?

[Mr. Morden]

Surely the point of a prediction is it must anticipate this to an extent? A prediction that is constantly changing isn't a prediction any more, it's just another tracker.

How is a model based on polls supposed to "know" what the polls will be in the future?  All it can do is say that, based on what the polls say at the moment, here's the probability of each candidate winning.  If the polls swing in a particular state from one candidate being up by 5 to a virtual tie, then there should be a big swing in the probabilities.

[Mr. Morden]

Right now, Huffington Post has chances for Clinton of winning FL at 97.1% and NC at 90.1%, which are higher chances than any of the other forecasting sites, including PEC.

By comparison, 538 shows chances for Clinton winning these two states at 50% each.

Perhaps the early/absentee vote numbers are a big reason for the differences?

As I said in the 538 thread, the models don't have any way to account for early voting, so no, that's not it.  I'm assuming that part of it is that the HuffPo model is slow to react to new polls, probably because the correlations between states are weaker.  So eventually, if the polls stayed like this, HuffPo would catch up to 538 in narrowing the gap.  Of course, there might not be an "eventually" because the election is just a few days away.

[mark_twain]

Right now, Huffington Post has chances for Clinton of winning FL at 97.1% and NC at 90.1%, which are higher chances than any of the other forecasting sites, including PEC.

By comparison, 538 shows chances for Clinton winning these two states at 50% each.

Perhaps the early/absentee vote numbers are a big reason for the differences?

As I said in the 538 thread, the models don't have any way to account for early voting, so no, that's not it.  I'm assuming that part of it is that the HuffPo model is slow to react to new polls, probably because the correlations between states are weaker.  So eventually, if the polls stayed like this, HuffPo would catch up to 538 in narrowing the gap.  Of course, there might not be an "eventually" because the election is just a few days away.

Never going to happen, even if Election Day weren't next Tuesday.

By now, Clinton has well recovered from the Comey news.

HuffPost's numbers for FL have moved slightly downward from 97% to 95%, due to the updates from recent polls.

The only explanation for this phenomenon is that HuffPost has special information not available to the other forecasters, specifically regarding early and absentee voting.

[Mr. Morden]

Right now, Huffington Post has chances for Clinton of winning FL at 97.1% and NC at 90.1%, which are higher chances than any of the other forecasting sites, including PEC.

By comparison, 538 shows chances for Clinton winning these two states at 50% each.

Perhaps the early/absentee vote numbers are a big reason for the differences?

As I said in the 538 thread, the models don't have any way to account for early voting, so no, that's not it.  I'm assuming that part of it is that the HuffPo model is slow to react to new polls, probably because the correlations between states are weaker.  So eventually, if the polls stayed like this, HuffPo would catch up to 538 in narrowing the gap.  Of course, there might not be an "eventually" because the election is just a few days away.

Never going to happen, even if Election Day weren't next Tuesday.

By now, Clinton has well recovered from the Comey news.

HuffPost's numbers for FL have moved slightly downward from 97% to 95%, due to the updates from recent polls.

The only explanation for this phenomenon is that HuffPost has special information not available to the other forecasters, specifically regarding early and absentee voting.

I’m sorry, but I think you’re absolutely wrong on this.  Not the point about Clinton rebounding in the polls.  That part may well be right.  What I think you’re wrong about is the idea that if a forecast model turns out to be more accurate, then it must be because it’s better rather than just more lucky, or that it must be because the data being put into the model is more complete, or that that data might include early voting numbers.

You have to distinguish between three different things:

1) What datasets do you include in your model?

2) How do you use those datasets in your model?

3) How important are factors that you don’t include in your model at all, and do they shift things in such a way to make your model look better or worse?

On #1, I have no reason to believe that any of the models is using something other than what they say they’re using.  For the most part, they’re just using publicly available polls.  I’m not aware of any of them using early voter data (except to the extent that early votes are themselves reflected in at least some of the polls), and I’m not sure how they would even use it if they wanted to.

On #2, yes, there are differences in how the models use the data.  As I said, apparently the 538 model has the states being much more heavily correlated with each other than the other models do, which means that from state polls alone, you can get significant shifts in the 538 model that wouldn’t show up in HuffPo.

I don’t know if they’re “right” to do it that way or not.  Maybe they are and maybe they’re not.  But my point was that they may well be right, yet unlucky.  For all we know, there was a real shift towards Trump in the polls and it wasn’t just noise, but then there is / will be another shift back towards Clinton just as we reach election day.  Or maybe there is no shift back in the polls, but the polls are all systematically underestimating Clinton because of ground game / failure to poll in Spanish / whatever.

In that case, the 538 model, which was sensitive to those polling shifts while other models weren’t, would turn out to be more wrong.  HuffPo would turn out to be more right, but it wouldn’t be because they built a better model or were using better data.  It would be because they were using a worse model, but their model’s failures ended up getting cancelled out by (#3) other factors that they weren’t even considering.  It would mean that they were more lucky than better.