As horrid as Facebook's intent typically is, and this offering certainly qualifies as such IMHO, "crowdsourced predictions and collective insights" has nothing to do with online voting.
What I find most interesting is the system for attributing online credibility to both experts and non-experts by history of prediction performance.
Prediction markets also typically use money to incentive people not to make garbage bets, but by having web credibility be the stakes, the conversation of bad incentives for bad events is a little more distant.
This is what I find most exciting, ever since I ran across the story of Superforecasters, who supposedly develop techniques to narrow down the probabilities of specific future events within very accurate bounds.
The point of black swan events is that you can't account for them. Prediction is irrelevant because you cannot compute the probability of an event that has never occurred before. That is why you care about robustness, prediction is inherently fragile.
I have read most of Taleb's books and I'm aware of his objections (and public beef with Tetlock).
Nonetheless, empirically some people are better than others at the task of assigning probabilities to events (as insanely hard as it is!) at a level above random, and it's a learnable skill that gets better with more people.
Yep, I worked in finance for years (investing capital) and I have bet on sports profitably for years. My point was actually nothing to do with superforecasting but superforecasting is obvious bullshit. Probabilities aren't computable in any consistent way, and the nature of these events mean that it is often very difficult to work out if someone's track record is actually real (example: the sample size you need for events that occur infrequently is, in practice, much larger than the sample size you need for events that occur more frequently). In the real world, all that matters is information. The idea that there is a group of some super-smart genius forecasters is a rather common mistake of academics...in the real world, the issue is not having information (i.e. the outcome of a vote is not a random event, that is why investment banks/hedge funds pay politicians for information). Btw, I think philosophers in the late 18th century understood this...suggesting that you can sit in your ivory tower and just reason through all these things should seem ludicrous.
Also, you cannot predict black swans. If you think you can, you don't understand what black swan is. The point of a black swan is that it is unpredictable because it has never occurred before.
> The idea that there is a group of some super-smart genius forecasters is a rather common mistake of academics
I'm pointing out is that there do exist some people who are more accurate than the others at assigning probabilities consistently, year after year. This is the real world phenomenon that needs explanation (though see below for caveats).
> Also, you cannot predict black swans
You don't have to exactly predict black swans to roughly estimate probabilities of events. As in, you don't need to enumerate exactly what could happen in the future, you do need to be aware of the possibility but some events are more likely than others.
Now, Taleb would say this is all nonsense because you can't meaningfully deduce a probability of a future event without knowing the denominator, that is look at all future possible worlds leading to outcome A as the numerator and (all future possible worlds leading to outcome B + worlds leading to A) as the denominator.
Which is strictly speaking very true and we are liable to fool themselves. Yet at the same time you can deduce that on aggregate people above the age of 80 have fewer years to live than people under 8... and if you were to predict how long someone would live based on their age, would you just throw your hands up and say it's impossible to know because of black swan events? Or might you look at actuarial tables and try to come up with a probability, especially if there's profit to be had in the form of selling insurance...
Taleb would counter that these predictions are fragile, that one day a black swan will come along and steamroll you when the next cosmic gamma ray event wipes out only the young and not the old, wiping out the past 200 years of gains.
Is that what you're referring to when you say that (super)forecasting doesn't work?
No, you can account for Black Swan events: Talib's whole company is built around the idea that they are more probable than people assume.
The Superforecaster approach absolutely should account for this, and to some degree does. You won't see anyone who has done superforecaster training saying things like "It is 99% likely X will win a two horse race election" before any votes have been cast.
I thought the real take away was that we (humans) have an inherently bad understanding of the true nature of randomness.
He suggests that a 'Black Swan' is still a 'Black Swan' (completely new, and totally unpredictable), but there are also classes of events that we 'think' we can ascribe a probability to them, and others where we actually can.
The problem is in conflating the two.
His investment strategy, as I understand it, is to capitalize on that confusion. Which is exceptionally hard work, in spite of a summary of Taleb's idea.
I'm somewhat of a fan of Taleb but I've never quite figured out how this works out in practice.
1. Taleb says the stock market systematically under-prices risk when in reality probabilities have fat tails, or high kurtosis.
2. Taleb took advantage of this by buying OTM puts which were being undervalued compared to the real risk. He's now "F you" rich, awesome.
(correct me if I'm wrong on these two assumptions)
The next question I have is, now that the cat's out of the bag, does this strategy still work? Is it possible that the market could over-value tail risk, making it a losing proposition to run this strategy (at least on certain stocks).
And if so, how do you know when that deep OTM put option is over or under priced?
I recall Taleb spent a lot of time talking about the problems but was light on actual details of how to make this strategy work out in practice (though it's been a few years since I've read some of his earlier books).
Hopefully you will see this a few days after the fact.
Markets, and people generally, don't work like that. The exact logic varies with strategy but, in this case, people like strategies that involve picking up quarters in front of a steamroller (i.e. selling deep OTM options).
These strategies sell exceptionally well with retail investors. French investment banks, in particular, are known for absolutely loving these strategies. And they blew up hundreds of thousands of Korean investors a couple of years ago with this stuff (and they did this in 2008 too...people really love these strategies and French banks really love blowing up retail investors).
A slightly more trenchant example is tech stocks. These are high momentum stocks, we know that high momentum is extremely cyclical and will eventually blow up (i.e. like option selling) but people do it because they think they will be able to jump out before it goes wrong (spoiler: they won't).
So yes, there is always a price at which something stops working but it is very unlikely we are at that point with this strategy. Implementation of this strategy varies. Empirica (or whatever they are called) apparently do something clever, I know someone doing this who just buys 10-15% OTM puts (and does some kind of simple backtesting/valuation)...but the basic idea is: what is the probability of the price falling this much, what do I gain if this hits, what do I lose if it doesn't. You can get very complex with modelling volatility, if you are trading volatility then that is relevant...for this strategy, you have no real idea when it is going to occur (the PM I know who was buying puts was doing it for three or four years before this happened, they still a lost a bit in the drop but, importantly, they didn't have to sell anything so have had a great year (up 10-15% with very little net exposure to markets).
The training that good judgement project gives you is a beautiful introduction to all sorts of evidence weighing, prediction, black swans, power dynamics of an outcome etc. Should be required for more professions and possibly democracies
So this is fortune teller gambling? Like people are trading and betting on people's predictions?
Facebook's version of betting on horseraces and other sporting events?
Am I wrong? I'm not sure exactly what this is other than you're trading in people's predictions.
Which pretty much sounds like when you go down to a bar and place bets on a boxing match or something....except with everything and it's run by facebook.
Like this sounds like a marketplace for people to put odds on everything then other people come place bets and it's all run by facebook.
Did I miss something here or interpret this wrong??
This seems ridiculous...it should be called gambling on global affairs by facebook..
But it is definitely true that this is different from betting generally: if I bet on a game, my bet doesn't change the outcome. In any social context, the probability on these markets (if it did take off) would affect behaviour.
Here's another way to look at it. Assume I risk massive downside if Candidate A gets elected, or if Bill B passes. I would like to buy "insurance" against these events occurring: I want to get a financial upside to cover my anticipated loss. So I'd put a bet on/option Candidate A winning, or on Bill B passing.
And then if Candidate A wins, I get a payout and use it to finance my moving abroad.
This doesn't use real money? Just worthless "points"?
If it doesn't use real money the predictions will be garbage.
"Why do you have to use real money?"
Prediction markets work best when players have some stake, however small, in the outcome. With play money, many players take risks they wouldn’t otherwise take or don’t attend to their holdings as carefully. Such markets may therefore have less research value than real money ones. Besides, we think real money is fun too.
It seems to use reputation instead, which might do decently with those who take it seriously. For example, the Good Judgement Project forecasting competition uses points and scores to incentivize: https://www.gjopen.com/faq
> Predictions are scored using Brier scores.[6] The top forecasters in GJP are "reportedly 30% better than intelligence officers with access to actual classified information
Yea it’s a forecasting competition but it does seem related in the way that if you were to take things seriously, step one would be to estimate the odds, step two is to buy the position that’s underpriced.
Presuming that you have to invest points and can't go too far in debt, it means people with a better track record at predicting have more weight to their predictions. Long term it seems like this selects for participants who are correctly attempting to weigh the options, and are good at it.
There seems to be a maximum amount of points any one person can invest in a prediction right now. There is a cap of a 10 share stake in any single prediction and a cap of 99 points per any share. This cap will prevent power players from placing too much weight on predictions.
Wait what? That ruins the ability of anyone successful to weight their predictions based on confidence (they'll just fully weight everything).
Will be interesting to see how this one plays out. I wonder what effect that will have on the market, it makes it much closer to a democracy than a betting market.
It’s legal for PredictIt. If FB could get some academic researchers on board I could see it happening but I guess it was more of a headache than what they wanted.
Points will work fine. The reason money works is partly the SKITG effect, but not entirely. In an iterated prediction market game, eventually reputation will accrue to those that make accurate forecasts. Much in the same way that money accrues to those skilled in business, or at forecasting in traditional financial markets. This iterated equilibrium should result in high quality predictions after a relatively short period of calibration.
I'm not convinced. If there's no direct profit, I can't imagine talented people are going to waste that much time on it.
Reputation is also not that great of a proxy, since you can game it. E.g. by using strategies that push all the risk into the tails. If it's money, you have a much stronger incentive to cover your ass against rare but disastrous blowups, since your entire livelihood is at stake (rather than just your ability to play the local game). If you're just trying to look good on Facebook... The risk/reward is different.
> I'm not convinced. If there's no direct profit, I can't imagine talented people are going to waste that much time on it.
Wikipedia, reddit, HN, and stackoverflow would beg to differ. Reputation drives people to expend incredible amounts of effort. In 2002 I would have had the same intuition you do. But i've seen reputation systems and social credit build too many incredible things to feel that way anymore.
> Reputation is also not that great of a proxy, it encourages strategies which push the risks into the tails. If it's money, you have a much stronger incentive to reduce your own tail risk since your entire livelihood is at stake, rather than just your ability to play the local game.
I do think there's a reasonable argument that the risk curves will be different, but I don't really see how that cashes out in an inaccurate overall system. It seems to me that as long as reputation gets allocated to those who are consistently correct, the system should work pretty well.
There is no StackOverflow for trading, or betting. Nobody gives away their edge for free, although they may give away their trades.
>I don't really see how that cashes out in an inaccurate overall system
Imagine a strategy that has a 50% chance of blowing you up completely over 10 years, but nets you 50% more profit if it works. The average return of this strategy is 25% less than a safe strategy (zero risk, zero gain), but it has a 50% chance of winning in this marketplace. If it were real money, nobody would go near it.
Ok, now imagine there's 100 people running strategies with that profile and 100 running the safe strategy. All of the top spots will be held by the suicidal strategy. This is an artificial example but you get the point. Real money forces you to avoid that risk profile as soon as you identify it.
> There is no StackOverflow for trading, or betting. Nobody gives away their edge for free, although they may give away their trades.
This isn't trading, though. It's forecasting. You could make that argument about the sub-category of economic forecasts, perhaps. But not about more general things, like presidential elections.
> Imagine a strategy that has a 50% chance of blowing you up completely over 10 years, but nets you 50% more profit if it works. The average return of this strategy is 25% less than a safe strategy (zero risk, zero gain), but it has a 50% chance of winning in this marketplace. If it were real money, nobody would go near it.
People go near real money strategies like that all the time. That's all WallStreetBets is, or the financial crisis, for that matter. I do take your point that this behavior may be more pronounced in a community like this, but I don't think it invalidates the model.
That's not my point. The point is people don't give away predictive edges for internet points (or reputation). There are many reasons why but the lack of a betting & trading StackOverflow is one symptom. It's just an example, but TBH it's not that relevant since this market wouldn't involve revealing edges.
>People go near real money strategies like that all the time.
Absolutely, and it's a problem with real funds. But:
- It's a lot rarer.
- There's a much stronger incentive for other people to quietly take the opposite sides of those bets, which normalises the market.
We'll see what effect it has, but I think this marketplace will just end up following the values on PredictIt.
The average answer/comment is pretty bad in most of the sites you listed. All of those use heavy ordering to work. Unlike say if I ask say kubernetes expert in my company some question, the chances of it being pure bad is very less compared to if I ask in stack overflow. Wikipedia has altogether different politics and incentive that it is not that open for most of the popular articles.
Right now there seems to be a cap of 99 tokens per person for each auction, which will clamp the effect rich players have on the weightings (to the point of it being much closer to an average). They may remove that in the future of course.
Not only is the average not that great but the fact that the lowest common denominator buries the actual experts on some of those platforms is maddening.
>Wikipedia, reddit, HN, and stackoverflow would beg to differ
Ha!
It's pretty obvious that on all those platforms the "people with tons of points" cohort is overwhelmingly made up of people who've gamed the system (which in pretty much all cases takes the form of "producing high volumes of low quality content that caters to the lowest common denominator of the local user demographics and doing so for a long time").
I agree. My eyes popped when I saw the headline as I was eager to participate. When I saw it didn’t involve real money I didn’t bother to read further.
You right, just like here, on HN, with this comment that I'm writing you all right now, hoping that it won't get downvoted, or that I won't get laughed at, because I'm smart dammit, and I have 1300 points that I don't want to lose.
But then there's so many others like me who don't give a damn about these points, mostly because they mean very less in this community. Yes, you could collect more points by writing smarter factual stuff but the same can also be done by writing stuff that appeals to a certain section of the people and these people have the loudest voices (mostly active on the site).
I'm new to HN though and I'm curious if someone can change my mind.
People do vastly different things when money is involved. For example, using 50 points to get another emoji in a chat app is ok, but over my dead body that I will pay a single cent to get it.
I’ve even seen that getting people to bet with real money and they will be exponentially more careful, diligent and down to earth even if it’s just lunch money.
For that you’d also have to be betting serious amounts of money. Take a trip to Vegas or options trading forums and you’ll see some people get very serious while others trade “for fun”
I've been using poly.market. Their upcoming upgrade won't have fees (allegedly). Right now it's expensive to place trades because it uses transactions on Ethereum.
Like the difference between playing poker with a table full of serious guys who want to win and playing poker "for fun" and no stakes with a group of friends half of which have never even played poker. Night and day experiences.
This seems like a strange thing for Facebook to build. I'm sure I'm missing some grand idea of why Facebook would make this. I just don't see how they get much out of this and how it fits into their mission.
Edit: I missed that this was coming out of NPE (New Product Experiments). Still I'm not sure why this would make sense for FB to run, but it's less odd knowing that it's an experiment.
I think it actually makes sense, potentially, as a way for FB to try and circumvent the publisher vs. platform debate.
If they can credential experts based on some metric that isn't based on their editorial discretion, they have more grounds to call themselves a "platform." Not saying I agree, but makes sense given their position.
economics and mechanism design is in the DNA of both Facebook and Google — core business of both is running auctions. that background tends to go with thinking prediction markets are cool.
Well, why try so hard to predict what people think, when you can just ask them and they will gladly tell you? As a bonus, you also get a motivation that can be exploited for much better targeting of, uhm, educational ads?
Data gathered from here is explicit, and this much more reliable than what you can implicitly infer from users' behavior.
This was my first thought. My guess is the team making the app are all about it's potential and fb are happy to give it a go because of the potential for improved data capture
Hi HN, I helped build Forecast. We're a small team within Facebook's New Product Experimentation group. We think prediction markets have the potential to create structured conversation spaces, to drive constructive discussion about what might happen in the world in a time when that's pretty hard to find. Please give the product a try and let us know if you have feedback!
I ran a project in this area for a few years, and saw significant issues. For example, for high-stakes predictions (eg elections) we saw significant gaming because people see it as an opportunity to move the media narrative.
We've seen this in the current US Presidential Elections, with big swings in betting markets for the presidential outcome itself, without the same swings in related outcomes (eg, you see a swing in the price for "the winning president is X" without the price for "X gets above 270 electoral votes changing at all")
Obviously this creates arbitrage opportunities, but the media narrative is the real issue. Have you considered this problem at all?
This is something we're thinking about continuously; it's unclear how central the prediction %s themselves should be. I am hopeful our core community can become more accurate over time, and get better and better at surfacing new information to make the market more efficient. We'll be evaluating this throughout...
That's good, but I suspect continual monitoring of headline markets ("who will win the US Election") is insufficient because FB is big enough to influence the result.
What are you going to do if (when) you detect manipulation? You won't be able to take it offline without the President using it as an example of "liberal bias", and you won't be able to ban users because of the same thing.
I know you won't do this, but I think you should experiment with some lower-stake markets first, and potentially use participation in those as a gate to the high stakes markets.
Gating could be a good idea; I think also leaning less on the aggregate/average and more on the positions from reliable participants in the market would be important once there's any kind of influence or scale to this product.
These are really really important points and the fact that Facebook is launching products without solid and convincing answers to these issues is a big worry.
This is a fascinating product and potentially of great value but the size and visibility of Facebook means that the potential for unintended consequences is BIG.
Hi Taylor, this is one of the cooler (probably coolest) consumer products to come out of facebook in a while. I was following it closely since the announcement a few months back.
My question is: I am sure you guys are familiar with crypto-based prediction markets like augur.net and prediqt.everipedia.org that use smart contracts and real money for predictions. Given that facebook is also building Libra and getting into blockchain, do you guys see that down the line Forecast could be something that moves to facebook's blockchain and/or you guys collaborate with crypto-based products like the above mentioned? That would be truly exciting since it would bring pure market dynamics to the predictions. I'm aware of the regulatory hurdles there, but thought I'd ask anyway. Great job! This is so fun to use.
Thanks! Right now we're focused more on building the community and usability of the product, so the market implementation itself is fairly simple. But we're following the various crypto implementations closely...
Community is always the most important aspect, agreed! Great to hear you guys are watching the crypto prediction markets closely, they are coming up with some novel features. I'm sure they'd all be interested in working with you guys at Facebook too given the users they could get through a partnership. Good luck!
Hey Taylor, cool product! But why can't you see the purchase prices of the arguments? As the price moves around over time you can try to look at the date and graph independently, but some indicator showing the movement since purchase on each of everyone's arguments would be a lot more useful! Cheers.
• In Chrome, the website log in system fails silently (and we get "FB.login() called when user is already connected" when retrying).
• The website is blank on Firefox.
• The principle is cool, but not Bayesian enough IMO. Points are expendable (especially when one can get them freely by creating a new account). Better use statistical priors estimating the likelihood that a given source outputs an accurate prediction. Give a bad prior to new accounts, to discourage creating new accounts to "clean the slate" of bad predictions.
• In the current state, this system, if it gets popular, can quickly have diplomatic, political and economic repercussions and consequent abuse that the website seems wholly unprepared for, based on my reading of the guidelines.
Hey Taylor, and here I was just wondering what NPE was and finding the thought that Facebook farmed out an app to another company amusing. Anyway, your website is trash, but the app is pretty nice and I'm enjoying it so far.
This doesn't appear to be a prediction market at all. There's no price signal or equivalent constrained property (and no - reputation isn't the same: You can't "stake $X of my reputation on this precise prediction").
It seems to be more like a clone of the Good Judgement Project.
Edit: Apparently there are shares. I can't tell because I'm not in the US/Canada and they don't appear mentioned anywhere.
I'll join the crowd and say, if this doesn't use real money, it's essentially worthless.
Even if people value a "reputation" the way they want HN or Reddit karma... people's behavior is ultimately extremely different. They make wilder bets "just for fun", because the rewards are highly asymmetrical: everybody cares if you're top of the leaderboard, nobody cares if you're in the bottom 95%, so the rational strategy is to make wild improbable bets and hope you get lucky, because there's lots of upside if you win and zero downside if you don't. (Or create 1,000+ different accounts to cover every combination of outcomes, and only publicly claim ownership of the one that wins the most.) All of this necessarily "pollutes" predictions making them ultimately worthless.
Prediction markets have to use real money people care about not losing. There is simply no way around that principle.
Poker players are well aware of this: advice on learning poker stresses how important it is to play for money as opposed to "fake money", because you and the other players will all play differently. (Play money is only when you're learning/practicing the basic mechanics of the game.)
Now of course, prediction markets have the unfortunate necessary side effect of people making money off of morally abhorrent events -- literally profiting off of a genocide or terrorist bombing -- when they successfully predict them. And a lot of people also see it as closer to gambling than investing. Which is why I can't ever imagine a large corporation like Facebook designing one -- it would be suicidal PR.
There's another discussion a similar kind above, but I think the Open Source community, Twitter, this place, Reddit, and even the likes of 4Chan, have proven that very able people will happily make exceptional contributions to various areas of knowledge, for a seemingly intangible incentive.
What's more, with Facebook attached, this will probably be the most prominent predication market yet known... Tbh, it would have been more on brand for Twitter to do this, but it's still very cool. I think the volume and kudos associated with the scale the brand name will bring, will make this one of the more successful products of it's type, in terms of results.
Oh I'm not disputing that people will contribute at all, in fact I'm positively sure they will.
The issue is that people's contributions will be highly skewed, for the reasons I listed.
Here's an analogy: imagine someone starts a separate stock market with play money, but asks everyone to invest "as if they would with real money". Do you really think that stock market will track the real stock market, even remotely?
Of course not. It'll be full of wild swings, people will make decisions for entertainment, nobody will pay attention to market fundamentals, because nothing is tethering it to reality.
A prediction market without real money suffers from the exact same problem.
I don't think open source or internet forums are a valuable analogy here, unfortunately -- pull requests are accepted or rejected by a maintainer, Wikipedia has very strict rules on what is accepted, and misinformation on Twitter and Reddit is rampant. A prediction market attempts to provide actual objective predictions without anyone in charge, which makes it totally different.
I have a small feature request after using this product.
When showing how many points another user has, such as next to a comment they made, it would be better to show their entire net worth of points (if they sold everything how many points would they have) instead of their available to spend points. I’m interested in knowing how successful their previous predictions have been, not in how much liquid capital they have.
Did not see this mentioned in any of the comments, but Metaculus [1] has been around for a while.
Do you directly compete with them? Nothing wrong in competing, I have a feeling that people are probably more _aggressive_ (in a positive sense) in their outlook when "Facebook profiles" are not on the line. Nevertheless, I think it is a great experiment.
I do imagine that at some point, if this product goes mainstream, journalists would start referring to these statistics for topics which really require correcting for exposure bias. For instance, exit polls. In places where Facebook penetration (or even internet penetration) is not-so-great, this will blindside decisions. Of course, this by no means will be Facebook's fault but I am curious to see how Facebook is proactively adjusting for this adverse outcome.
An exciting day to launch a new prediction market, when Trump announces he's caught Covid and Betfair has to suspend their presidential election odds market....
It's odd, isn't it, many US posters here seem to think prediction markets are new and exciting, yet seem unaware of non-US sites like that Betfair, which has been hosting real-money betting markets on politics for well over a decade. Their 'next US president' market already has over £100 million matched.
(And Mike Pence betting on a 999/1 -> 39/1 plunge, presumably an over-reaction to the COVID news prior to the market being suspended!)
135 comments
[ 2.0 ms ] story [ 200 ms ] threadhttps://www.legalsportsreport.com/37949/2020-presidential-el...
https://abcnews.go.com/Politics/story?id=122622&page=1
https://www.cnn.com/2003/ALLPOLITICS/07/29/terror.market/
Also, it is a form of engagement for Facebook users, which is valuable in itself.
It came from the New Product Experimentation team.
Prediction markets also typically use money to incentive people not to make garbage bets, but by having web credibility be the stakes, the conversation of bad incentives for bad events is a little more distant.
https://en.wikipedia.org/wiki/Superforecasting:_The_Art_and_...
A lot of work has to be done on overcoming cognitive biases and accounting for black swan events in order to attain high level performance.
Nonetheless, empirically some people are better than others at the task of assigning probabilities to events (as insanely hard as it is!) at a level above random, and it's a learnable skill that gets better with more people.
https://commoncog.com/blog/forecasting-under-uncertainty/
Also, you cannot predict black swans. If you think you can, you don't understand what black swan is. The point of a black swan is that it is unpredictable because it has never occurred before.
I'm pointing out is that there do exist some people who are more accurate than the others at assigning probabilities consistently, year after year. This is the real world phenomenon that needs explanation (though see below for caveats).
> Also, you cannot predict black swans
You don't have to exactly predict black swans to roughly estimate probabilities of events. As in, you don't need to enumerate exactly what could happen in the future, you do need to be aware of the possibility but some events are more likely than others.
Now, Taleb would say this is all nonsense because you can't meaningfully deduce a probability of a future event without knowing the denominator, that is look at all future possible worlds leading to outcome A as the numerator and (all future possible worlds leading to outcome B + worlds leading to A) as the denominator.
Which is strictly speaking very true and we are liable to fool themselves. Yet at the same time you can deduce that on aggregate people above the age of 80 have fewer years to live than people under 8... and if you were to predict how long someone would live based on their age, would you just throw your hands up and say it's impossible to know because of black swan events? Or might you look at actuarial tables and try to come up with a probability, especially if there's profit to be had in the form of selling insurance...
Taleb would counter that these predictions are fragile, that one day a black swan will come along and steamroll you when the next cosmic gamma ray event wipes out only the young and not the old, wiping out the past 200 years of gains.
Is that what you're referring to when you say that (super)forecasting doesn't work?
The Superforecaster approach absolutely should account for this, and to some degree does. You won't see anyone who has done superforecaster training saying things like "It is 99% likely X will win a two horse race election" before any votes have been cast.
The problem is in conflating the two.
His investment strategy, as I understand it, is to capitalize on that confusion. Which is exceptionally hard work, in spite of a summary of Taleb's idea.
1. Taleb says the stock market systematically under-prices risk when in reality probabilities have fat tails, or high kurtosis.
2. Taleb took advantage of this by buying OTM puts which were being undervalued compared to the real risk. He's now "F you" rich, awesome.
(correct me if I'm wrong on these two assumptions)
The next question I have is, now that the cat's out of the bag, does this strategy still work? Is it possible that the market could over-value tail risk, making it a losing proposition to run this strategy (at least on certain stocks).
And if so, how do you know when that deep OTM put option is over or under priced?
I recall Taleb spent a lot of time talking about the problems but was light on actual details of how to make this strategy work out in practice (though it's been a few years since I've read some of his earlier books).
Markets, and people generally, don't work like that. The exact logic varies with strategy but, in this case, people like strategies that involve picking up quarters in front of a steamroller (i.e. selling deep OTM options).
These strategies sell exceptionally well with retail investors. French investment banks, in particular, are known for absolutely loving these strategies. And they blew up hundreds of thousands of Korean investors a couple of years ago with this stuff (and they did this in 2008 too...people really love these strategies and French banks really love blowing up retail investors).
A slightly more trenchant example is tech stocks. These are high momentum stocks, we know that high momentum is extremely cyclical and will eventually blow up (i.e. like option selling) but people do it because they think they will be able to jump out before it goes wrong (spoiler: they won't).
So yes, there is always a price at which something stops working but it is very unlikely we are at that point with this strategy. Implementation of this strategy varies. Empirica (or whatever they are called) apparently do something clever, I know someone doing this who just buys 10-15% OTM puts (and does some kind of simple backtesting/valuation)...but the basic idea is: what is the probability of the price falling this much, what do I gain if this hits, what do I lose if it doesn't. You can get very complex with modelling volatility, if you are trading volatility then that is relevant...for this strategy, you have no real idea when it is going to occur (the PM I know who was buying puts was doing it for three or four years before this happened, they still a lost a bit in the drop but, importantly, they didn't have to sell anything so have had a great year (up 10-15% with very little net exposure to markets).
Facebook's version of betting on horseraces and other sporting events?
Am I wrong? I'm not sure exactly what this is other than you're trading in people's predictions.
Which pretty much sounds like when you go down to a bar and place bets on a boxing match or something....except with everything and it's run by facebook.
Like this sounds like a marketplace for people to put odds on everything then other people come place bets and it's all run by facebook.
Did I miss something here or interpret this wrong??
This seems ridiculous...it should be called gambling on global affairs by facebook..
A popular one is PredictIt: https://en.wikipedia.org/wiki/PredictIt
But in the end...call it what you want...it's gambling..
>skilled player can make tons of money by exploiting mispriced contracts
Except, unlike in most poker hands, those have far reaching global consequences
But it is definitely true that this is different from betting generally: if I bet on a game, my bet doesn't change the outcome. In any social context, the probability on these markets (if it did take off) would affect behaviour.
And then if Candidate A wins, I get a payout and use it to finance my moving abroad.
If it doesn't use real money the predictions will be garbage.
"Why do you have to use real money?"
Prediction markets work best when players have some stake, however small, in the outcome. With play money, many players take risks they wouldn’t otherwise take or don’t attend to their holdings as carefully. Such markets may therefore have less research value than real money ones. Besides, we think real money is fun too.
https://www.predictit.org/support/faq
https://en.wikipedia.org/wiki/The_Good_Judgment_Project
> Predictions are scored using Brier scores.[6] The top forecasters in GJP are "reportedly 30% better than intelligence officers with access to actual classified information
You can't bet more of your points in GJP if you are really sure about an answer, hence it isn't a prediction market.
Will be interesting to see how this one plays out. I wonder what effect that will have on the market, it makes it much closer to a democracy than a betting market.
Reputation is also not that great of a proxy, since you can game it. E.g. by using strategies that push all the risk into the tails. If it's money, you have a much stronger incentive to cover your ass against rare but disastrous blowups, since your entire livelihood is at stake (rather than just your ability to play the local game). If you're just trying to look good on Facebook... The risk/reward is different.
Wikipedia, reddit, HN, and stackoverflow would beg to differ. Reputation drives people to expend incredible amounts of effort. In 2002 I would have had the same intuition you do. But i've seen reputation systems and social credit build too many incredible things to feel that way anymore.
> Reputation is also not that great of a proxy, it encourages strategies which push the risks into the tails. If it's money, you have a much stronger incentive to reduce your own tail risk since your entire livelihood is at stake, rather than just your ability to play the local game.
I do think there's a reasonable argument that the risk curves will be different, but I don't really see how that cashes out in an inaccurate overall system. It seems to me that as long as reputation gets allocated to those who are consistently correct, the system should work pretty well.
>I don't really see how that cashes out in an inaccurate overall system
Imagine a strategy that has a 50% chance of blowing you up completely over 10 years, but nets you 50% more profit if it works. The average return of this strategy is 25% less than a safe strategy (zero risk, zero gain), but it has a 50% chance of winning in this marketplace. If it were real money, nobody would go near it.
Ok, now imagine there's 100 people running strategies with that profile and 100 running the safe strategy. All of the top spots will be held by the suicidal strategy. This is an artificial example but you get the point. Real money forces you to avoid that risk profile as soon as you identify it.
This isn't trading, though. It's forecasting. You could make that argument about the sub-category of economic forecasts, perhaps. But not about more general things, like presidential elections.
> Imagine a strategy that has a 50% chance of blowing you up completely over 10 years, but nets you 50% more profit if it works. The average return of this strategy is 25% less than a safe strategy (zero risk, zero gain), but it has a 50% chance of winning in this marketplace. If it were real money, nobody would go near it.
People go near real money strategies like that all the time. That's all WallStreetBets is, or the financial crisis, for that matter. I do take your point that this behavior may be more pronounced in a community like this, but I don't think it invalidates the model.
That's not my point. The point is people don't give away predictive edges for internet points (or reputation). There are many reasons why but the lack of a betting & trading StackOverflow is one symptom. It's just an example, but TBH it's not that relevant since this market wouldn't involve revealing edges.
>People go near real money strategies like that all the time.
Absolutely, and it's a problem with real funds. But:
- It's a lot rarer.
- There's a much stronger incentive for other people to quietly take the opposite sides of those bets, which normalises the market.
We'll see what effect it has, but I think this marketplace will just end up following the values on PredictIt.
https://news.ycombinator.com/item?id=24658738
Ha!
It's pretty obvious that on all those platforms the "people with tons of points" cohort is overwhelmingly made up of people who've gamed the system (which in pretty much all cases takes the form of "producing high volumes of low quality content that caters to the lowest common denominator of the local user demographics and doing so for a long time").
I'm new to HN though and I'm curious if someone can change my mind.
I’ve even seen that getting people to bet with real money and they will be exponentially more careful, diligent and down to earth even if it’s just lunch money.
That doesn't give it better predictive powers - and maybe the opposite.
Uprating people with a historical record of accuracy is likely to be at least better than wealth as a marker.
Edit: I missed that this was coming out of NPE (New Product Experiments). Still I'm not sure why this would make sense for FB to run, but it's less odd knowing that it's an experiment.
If they can credential experts based on some metric that isn't based on their editorial discretion, they have more grounds to call themselves a "platform." Not saying I agree, but makes sense given their position.
https://www.youtube.com/watch?v=jRUA0o-rJjU
Data gathered from here is explicit, and this much more reliable than what you can implicitly infer from users' behavior.
Why the decision to launch for iOS only?
I ran a project in this area for a few years, and saw significant issues. For example, for high-stakes predictions (eg elections) we saw significant gaming because people see it as an opportunity to move the media narrative.
We've seen this in the current US Presidential Elections, with big swings in betting markets for the presidential outcome itself, without the same swings in related outcomes (eg, you see a swing in the price for "the winning president is X" without the price for "X gets above 270 electoral votes changing at all")
Obviously this creates arbitrage opportunities, but the media narrative is the real issue. Have you considered this problem at all?
Edit: Here's an account of manipulation of betting markets during the Obama/Romney election: https://rajivsethi.blogspot.com/2013/09/the-romney-whale.htm...
What are you going to do if (when) you detect manipulation? You won't be able to take it offline without the President using it as an example of "liberal bias", and you won't be able to ban users because of the same thing.
I know you won't do this, but I think you should experiment with some lower-stake markets first, and potentially use participation in those as a gate to the high stakes markets.
This is a fascinating product and potentially of great value but the size and visibility of Facebook means that the potential for unintended consequences is BIG.
My question is: I am sure you guys are familiar with crypto-based prediction markets like augur.net and prediqt.everipedia.org that use smart contracts and real money for predictions. Given that facebook is also building Libra and getting into blockchain, do you guys see that down the line Forecast could be something that moves to facebook's blockchain and/or you guys collaborate with crypto-based products like the above mentioned? That would be truly exciting since it would bring pure market dynamics to the predictions. I'm aware of the regulatory hurdles there, but thought I'd ask anyway. Great job! This is so fun to use.
What technology is used for the mobile apps?
Has there been any risk assessment on this feature? If so, any chance this is made public?
• In Chrome, the website log in system fails silently (and we get "FB.login() called when user is already connected" when retrying).
• The website is blank on Firefox.
• The principle is cool, but not Bayesian enough IMO. Points are expendable (especially when one can get them freely by creating a new account). Better use statistical priors estimating the likelihood that a given source outputs an accurate prediction. Give a bad prior to new accounts, to discourage creating new accounts to "clean the slate" of bad predictions.
• In the current state, this system, if it gets popular, can quickly have diplomatic, political and economic repercussions and consequent abuse that the website seems wholly unprepared for, based on my reading of the guidelines.
Do you have Firefox FB containers enabled? I had the same issue, but when I opened the website inside a FB container it works.
It seems to be more like a clone of the Good Judgement Project.
Edit: Apparently there are shares. I can't tell because I'm not in the US/Canada and they don't appear mentioned anywhere.
Even if people value a "reputation" the way they want HN or Reddit karma... people's behavior is ultimately extremely different. They make wilder bets "just for fun", because the rewards are highly asymmetrical: everybody cares if you're top of the leaderboard, nobody cares if you're in the bottom 95%, so the rational strategy is to make wild improbable bets and hope you get lucky, because there's lots of upside if you win and zero downside if you don't. (Or create 1,000+ different accounts to cover every combination of outcomes, and only publicly claim ownership of the one that wins the most.) All of this necessarily "pollutes" predictions making them ultimately worthless.
Prediction markets have to use real money people care about not losing. There is simply no way around that principle.
Poker players are well aware of this: advice on learning poker stresses how important it is to play for money as opposed to "fake money", because you and the other players will all play differently. (Play money is only when you're learning/practicing the basic mechanics of the game.)
Now of course, prediction markets have the unfortunate necessary side effect of people making money off of morally abhorrent events -- literally profiting off of a genocide or terrorist bombing -- when they successfully predict them. And a lot of people also see it as closer to gambling than investing. Which is why I can't ever imagine a large corporation like Facebook designing one -- it would be suicidal PR.
What's more, with Facebook attached, this will probably be the most prominent predication market yet known... Tbh, it would have been more on brand for Twitter to do this, but it's still very cool. I think the volume and kudos associated with the scale the brand name will bring, will make this one of the more successful products of it's type, in terms of results.
The issue is that people's contributions will be highly skewed, for the reasons I listed.
Here's an analogy: imagine someone starts a separate stock market with play money, but asks everyone to invest "as if they would with real money". Do you really think that stock market will track the real stock market, even remotely?
Of course not. It'll be full of wild swings, people will make decisions for entertainment, nobody will pay attention to market fundamentals, because nothing is tethering it to reality.
A prediction market without real money suffers from the exact same problem.
I don't think open source or internet forums are a valuable analogy here, unfortunately -- pull requests are accepted or rejected by a maintainer, Wikipedia has very strict rules on what is accepted, and misinformation on Twitter and Reddit is rampant. A prediction market attempts to provide actual objective predictions without anyone in charge, which makes it totally different.
When showing how many points another user has, such as next to a comment they made, it would be better to show their entire net worth of points (if they sold everything how many points would they have) instead of their available to spend points. I’m interested in knowing how successful their previous predictions have been, not in how much liquid capital they have.
Do you directly compete with them? Nothing wrong in competing, I have a feeling that people are probably more _aggressive_ (in a positive sense) in their outlook when "Facebook profiles" are not on the line. Nevertheless, I think it is a great experiment.
I do imagine that at some point, if this product goes mainstream, journalists would start referring to these statistics for topics which really require correcting for exposure bias. For instance, exit polls. In places where Facebook penetration (or even internet penetration) is not-so-great, this will blindside decisions. Of course, this by no means will be Facebook's fault but I am curious to see how Facebook is proactively adjusting for this adverse outcome.
[1]: https://www.metaculus.com
(And Mike Pence betting on a 999/1 -> 39/1 plunge, presumably an over-reaction to the COVID news prior to the market being suspended!)
[0] https://smarkets.com
[1] https://www.predictit.org/
https://hunches.app/posts/2020-dem-nominee
[0]: https://asindu.drileba.capital/2020/06/the-wisdom-of-rationa...