3 ms·

The implied probability of Brexit a few days before was around 25%. The implied probability of Trump beating Clinton a few days before was around 20%. Neither

by neilkk 3y ago

The implied probability of Brexit a few days before was around 25%. The implied probability of Trump beating Clinton a few days before was around 20%. Neither the time value of money nor the cost of trading explain these dislocations.

eli 3y ago
The polls implied about a 20% chance of Trump beating Clinton too. I guess I'm not sure what you're getting at. That seems like the perfect example of not a sure thing.
blitzar 3y ago
> The implied probability of Brexit a few days before was around 25%. It was one of the 25 days out of 100 not one of the 75 days out of 100. > these dislocations The market puts the chance of flipping a coin and getting heads at 50%, when it comes up tails it doesn't mean that the market was irrationally pricing heads.
n2d4 3y ago
If anything then that means that they work! Situations with a 25% chance do occur sometimes! If prediction markets show 75% odds and are right 3/4 times, they are doing their job well - if they show 75% odds and right 4/4 times, they are doing their job badly! Meanwhile, both polls and analysts got those elections completely wrong, much more so than prediction markets. (IIRC, 538 was the only outlet that gave Trump a >20% chance to win, and they got a lot of bad words for their prediction pre-election.) Unfortunately, people keep falling for this. You can always cherrypick elections from 7 years ago as "evidence" against prediction markets, but if you look at the bigger picture, you will see that they get a vast majority of elections right. But when they don't, even just once in a decade, they get all the blame. There's plenty of research on this: https://researchdmr.com/RothschildPOQ2009.pdf https://researchdmr.com/RothschildPOQ2009.pdf
- MostlyStable 3y ago
  538 got screwed in both directions: before the election they were derided for having Trump too high. After the election they were lambasted for "only" giving him 20-30% chance.
  - cosmojg 3y ago
    I'm forever puzzled by the lack of rigorous coursework in probability and statistics in the American public school system, at all levels. I mean, heck, it's pretty much the only mathematical subject for which the question, "When are we ever going to use this in real life?" is trivially answered! And all the other more traditional math subjects (algebra, calculus, etc.) fall right out of it anyway. It's the perfect foundation. Honestly, why hasn't probability-centric mathematical pedagogy taken the world by storm?
nl 3y ago
> The implied probability of Trump beating Clinton a few days before was around 20%. Unclear if you think this is incorrect? Obviously it happened, but most reasonable prediction methodologies put it between 20 and 30%. Notably 538 and Superforecasters (via The Good Judgement Project at the time) both had it in that range (as did prediction markets). There were some outliers - in particular some newspaper(?) had a statistical methodology that gave Clinton a 95%+ chance. In a two horse race that is obviously wrong! I was working in the forecasting field at the time, and that seemed about right. It was an unlikely but not impossible outcome.
blackshaw 3y ago
So? 25% doesn't mean it can't happen. I wouldn't cross the road if there was a 25% of being hit by a car. The real way to measure predictive accuracy is to find _all_ the times they gave something a 25% chance of happening. If the predictions are accurate then roughly 25% of those things should have happened.
- nojs 3y ago
  If I am asked to predict whether it will rain every day for a year, and every day I blindly predict 30% because on average it rains 30% of days in this location, am I correct? Maybe, but I’m not a very useful forecaster. What you should optimise instead is something like log loss between the given probability and the true outcome (0 or 1). That way you’re rewarded not only for being right, but for being confident and right.

by neilkk 3y ago

eli 3y ago

The polls implied about a 20% chance of Trump beating Clinton too. I guess I'm not sure what you're getting at. That seems like the perfect example of not a sure thing.

blitzar 3y ago

> The implied probability of Brexit a few days before was around 25%. It was one of the 25 days out of 100 not one of the 75 days out of 100. > these dislocations The market puts the chance of flipping a coin and getting heads at 50%, when it comes up tails it doesn't mean that the market was irrationally pricing heads.

n2d4 3y ago

If anything then that means that they work! Situations with a 25% chance do occur sometimes! If prediction markets show 75% odds and are right 3/4 times, they are doing their job well - if they show 75% odds and right 4/4 times, they are doing their job badly! Meanwhile, both polls and analysts got those elections completely wrong, much more so than prediction markets. (IIRC, 538 was the only outlet that gave Trump a >20% chance to win, and they got a lot of bad words for their prediction pre-election.) Unfortunately, people keep falling for this. You can always cherrypick elections from 7 years ago as "evidence" against prediction markets, but if you look at the bigger picture, you will see that they get a vast majority of elections right. But when they don't, even just once in a decade, they get all the blame. There's plenty of research on this: https://researchdmr.com/RothschildPOQ2009.pdf https://researchdmr.com/RothschildPOQ2009.pdf

MostlyStable 3y ago
538 got screwed in both directions: before the election they were derided for having Trump too high. After the election they were lambasted for "only" giving him 20-30% chance.
- cosmojg 3y ago
  I'm forever puzzled by the lack of rigorous coursework in probability and statistics in the American public school system, at all levels. I mean, heck, it's pretty much the only mathematical subject for which the question, "When are we ever going to use this in real life?" is trivially answered! And all the other more traditional math subjects (algebra, calculus, etc.) fall right out of it anyway. It's the perfect foundation. Honestly, why hasn't probability-centric mathematical pedagogy taken the world by storm?

nl 3y ago

> The implied probability of Trump beating Clinton a few days before was around 20%. Unclear if you think this is incorrect? Obviously it happened, but most reasonable prediction methodologies put it between 20 and 30%. Notably 538 and Superforecasters (via The Good Judgement Project at the time) both had it in that range (as did prediction markets). There were some outliers - in particular some newspaper(?) had a statistical methodology that gave Clinton a 95%+ chance. In a two horse race that is obviously wrong! I was working in the forecasting field at the time, and that seemed about right. It was an unlikely but not impossible outcome.

blackshaw 3y ago

So? 25% doesn't mean it can't happen. I wouldn't cross the road if there was a 25% of being hit by a car. The real way to measure predictive accuracy is to find _all_ the times they gave something a 25% chance of happening. If the predictions are accurate then roughly 25% of those things should have happened.

nojs 3y ago
If I am asked to predict whether it will rain every day for a year, and every day I blindly predict 30% because on average it rains 30% of days in this location, am I correct? Maybe, but I’m not a very useful forecaster. What you should optimise instead is something like log loss between the given probability and the true outcome (0 or 1). That way you’re rewarded not only for being right, but for being confident and right.