…
Seems to be more of a polling issue than anything else. RCP has Indiana at Trump +10 and he’s currently +17 and Clinton +7 (currently Sanders +5)
538 doesn’t do their own polling, they just try to analyze and predict based off other polls.
So, they predicted a 10% chance of a Sanders win. 10% is not some unfathomably small chance. It’s a little less than winning three coin tosses in a row.
You’re falling into the media’s trap of seeing everything as binary. That everything is either a win or a loss, and every prediction is either right or wrong.
They don’t just pull numbers out of their ass, if there is a problem it is with the polling.
First of all I think the only other primary state the polls (and thus they) were way off on was MI. Spot checking a few: Ohio their average prediction was Clinton +11, result was Clinton +13.8; PA prediction Clinton +18, result +12; NY prediction +15, result +16; IL prediction Clinton +7, result +2; Texas prediction +32, result +32; SC prediction Clinton +38, result +47; VA prediction Clinton +28, result +29 … overall not so bad.
Second, on what do you base your belief that polls (and thus poll aggregation) have done much better in primaries in past years? My understanding is that primaries, especially open ones, are always tough predictions. The ability of pollsters to determine who is actually a likely voter in one or the other side is much harder than in a general election. And caucuses? Faggeddaboutit.
I think you are comparing solid performance in general election races in the past with primary prediction performance.
And I say this as one who has been on record arguing that polling is in tough times and who had been expecting epic fails for the aggregators as garbage in will produce garbage out.
They have done much better than I was expecting them to do.
this may be the wrong thread, but several news outlets have ted suspending hgis campaign. Go Blues!
I do not know if I can find past predictions in this primary, but this is not the first time 538 has been wildly inaccurate in this primary season.
You haven’t even shown them to be “wildly inaccurate” this time. I don’t know how else to say it. They said Sanders might win! They do not predict and never have predicted a single exact result.
If they predicted all 57ish contests with the same level of confidence, and were correct in that assessment, 5 or 6 would go the “wrong” way. And if they didn’t, that would be a flaw in the model.
It’s easy. On the primary prediction page they have a scroll down that includes past races that you can click to. It’s where I picked my several at random from poling plus. Caucuses I did not check but suspect that they were more off as caucus poll data is for shit. Roughing it my sense is that they were within their reported +/- 1 SD range something like 2/3s of the time. Which is, uh, what they should be doing. If they are not out of that range roughly one out three times then their model is off.
The questions that should matter:
Which has been more accurate - just the RCP rolling average, 538’s regression enhanced “polling only” model, or 538’s further enhanced secret sauce “polling plus” approach?
How has polling in general done predicting primaries this year compared to previous cycles with as prolonged of a season? Is polling itself broken or doing as well as always?
To the latter I’ve already cited Wang’s analysis that polling has done surprisingly well in aggregate this cycle.
If someone wants to crunch some data on a compare contrast of RCP rolling average and the two 538 approaches, have at it. I’d be interested.
If someone wants to crunch some past cycles polling performances, great!
This is off base for a couple of reasons.
-
This is the first time they’ve done primary forecasts. There is no past to compare to.
-
The polls only model has missed picking the winner exactly twice on the Dem side and three times on the Republican side. That is pretty good.
Tried to do another quick check going through 538’s poll’s plus … there have been 11 races that they gave the odds of a certain candidate winning as being roughly 90% (86 to 94% was what I used). In those they were wrong 1 out 11 times (Indiana). Pretty much exactly where they should be.
I didn’t count the number of the races higher than that, there were quite a few, all were correct except the one epic fail of Michigan.
So really, what “wildly inaccurate” are you looking at? That they underestimated caucus margins victories a few times? That they gave 80% confidence intervals of anything from 10 points one way to 10 points the other in Oklahoma with an average prediction essentially tied and it ended up 10 points one way (to Sanders)?