1/ The key to understanding "news" is to think about probabilities and incentives. A thread... πππ
2/ Let's say you're trying to estimate the chances of some future event happening.
As an example, let's take P(Trump runs for President in 2024).
As an example, let's take P(Trump runs for President in 2024).
3/ So you read the news about such an event. CNN, FoxNews, Politico, whatever. And it feels like a steady drip of "news" about tiny events which could affect that probability.
So how do we integrate all these pieces of information into a prediction?
So how do we integrate all these pieces of information into a prediction?
4/ Naively here's what you should do if you're a good Bayesian:
- Start with some baseline prior probability.
- For every news story you see, update that probability based on the content and importance of the news.
Doing this well gets you the best possible estimate.
- Start with some baseline prior probability.
- For every news story you see, update that probability based on the content and importance of the news.
Doing this well gets you the best possible estimate.
5/ But, while this isn't exactly *wrong*, it's close enough to wrong. Reason:
You're ignoring the process by which news stories are generated!!!
You're ignoring the process by which news stories are generated!!!
6/ The model of "news happens, media reports on it" is completely incorrect. There's way more news and pseudo-news that happens than media outlets have time to report on.
They perform an editing function by selecting what to report on.
They perform an editing function by selecting what to report on.
7/ So far this is all obvious, and it's also obvious that they select based on perceived interest (clicks, viewers, etc). But that doesn't help us yet.
The key question: "How do (perceived) probabilities affect story selection?"
The key question: "How do (perceived) probabilities affect story selection?"
8/ If you're a good Bayesian, you recognizes this as precisely Bayes Theorem. And here's what I'm hereby calling Lebron's Theory Of Media Probabilities:
"Media companies select stories which maximize entropy."
"Media companies select stories which maximize entropy."
9/ If everyone agrees something is going to happen, it's no longer news. There aren't any stories about the chances the sun won't come up tomorrow. If it didn't, that would be HUGE news.
But P(no sun tomorrow) = 0%.
Not news.
But P(no sun tomorrow) = 0%.
Not news.
10/ If you recall your 2nd year information theory for binary outcomes, entropy is maximized when P(true) = 50%.
For any binary outcome event, media are incentivized to report on stories which move the perceived probability to 50%.
That maximizes uncertainty and interest.
For any binary outcome event, media are incentivized to report on stories which move the perceived probability to 50%.
That maximizes uncertainty and interest.
11/ Will Trump run for President in 2024?
My read is that the probability is perceived to be under 50%, which is why news organizations are falling over themselves to find any indication he will. The more stories they report that he will, the more the prob moves closer to 50%.
My read is that the probability is perceived to be under 50%, which is why news organizations are falling over themselves to find any indication he will. The more stories they report that he will, the more the prob moves closer to 50%.
12/ Right now it seems likely Biden will run again. Over 50%.
So any piece of news that indicates Biden won't run is going to get a big push. Moving perceived probability closer to 50%.
So any piece of news that indicates Biden won't run is going to get a big push. Moving perceived probability closer to 50%.
13/ So when you're estimating probabilities, you have to back out the selection bias of the stories you're seeing. You have to undo the squeeze towards 50%.
They don't care about providing *accurate* probabilities. It's just not in their set of incentives.
They don't care about providing *accurate* probabilities. It's just not in their set of incentives.
14/ But if you're a trader, or otherwise in the business of making accurate predictions, you need to back out their incentives.
That way you can make good use of the info they provide.
That way you can make good use of the info they provide.
15/ This also generalizes to non-binary outcomes. After all, entropy is a property of any probability distribution.
TL;DR: News is an entropy-maximizing process.
/END
TL;DR: News is an entropy-maximizing process.
/END
β’ β’ β’
Missing some Tweet in this thread? You can try to
force a refresh
γ
