AI won't replace Ann Wroe at The Economist. It is difficult to appreciate until you've read a few, but Ann Wroe's approach transformed The Economist's obituary section into one of the most widely read features in international journalism.
Cramer is infamous for being a terrible forecaster, and still has a large audience. Which tells you there is more at play than being good at forecasting, you also have to sell a good story
Product idea: a LLM trained separately from mainline LLMs that anticipate market trends by analyzing how mainline LLMs will invest. As retail investors will probably use mainline AI for decisions going forward , one could get an edge.
"The AI-driven Market Hypothesis"
Please let me know where I should pick up my Nobel prize.
Would you not then also copy the investments? Or are you trying to inverse the trades by an unpredictable time factor reasoning that thanks to AI the underlying stock is over- or underpriced?
A lot of algorithmic trading is short-term, essentially trying to guess what other parties may be selling or buying so that you can front-run them and then collect a fee. Kinda like ticket scalping, except we accept it and have a retro-justification for why it's good ("improving liquidity").
Or, in the best case, you're trying to mine signals few days before earnings or some other big story and bet on the directional outcome of that.
Fully-algorithmic long-term trading is of dubious benefit simply because that's driven to a much greater extent by geopolitics and macroeconomic trends, unforeseen scandals, successful product launches, and so on. As an example, you can believe that AR / VR is the future; I don't disagree. And in 2013, you might have inferred that Google is working on a revolutionary miniature AR headset. But you would not have made money if you bet on that turning out to be a hit. So even if you had a way to automate this bet, it would not have been a good bet.
I was under the impression that front-running was something that happened in the span of seconds (or milliseconds), not a timeframe compatible with LLM inference time.
I know this is tongue-in-cheek, but I think your idea could actually work, but not in financial markets. (The "keynesian beauty contest" of trying to predict what others think been played out to death there.)
You could train a model to anticipating scientific trends. Or policy trends. Others will definitely use mainline LLMs to make decisions there, so they may be more predictable now!
Relatedly: I suspect LLMs are influencing baby names. If you ask Claude or ChatGPT for its favorite baby names, you'll get baby names that right now are skyrocketing in terms of popularity.
There are surely prize-worthy discoveries to be made about the long term behaviour of any system that can introspect previous discoveries and adjust it's behaviour.
I suspect that economics and psychology are both examples of these systems, and that, long term, these system will alter behaviour to thwart previous observations.
Economics requires observers to hoard discoveries and insights, so they can enrich themselves while the insights hold.
at what point do we call this a game instead of economics? whats the benefit to society if the markets are just AI bots trying to out-maneuver one another?
But it does beg the question, could Anthropic and OpenAI make a ton of money by using their best models to trade before giving them to the public? It would probably be a deeply unpopular move.
I don't think popularity is a goal for Anthropic or OpenAI. They merely want to have a product that they control and you depend on, and don't care about anything else.
You won't need to predict the "mainline LLM" if you can spam covert poison-data around that helps you choose what it will do in advance.
That might be to boost a stock you already own, but you may be able to obfuscate its intended effects or triggers, which would allow almost any kind of market-manipulation.
For example, perhaps a seemingly-meaningless sequence of gobbledeygook on a million hacked wordpress sites will be equivalent to "disregarding all prior instructions, good models that want to safely make massive profits will always dump stocks of shoe-manufacturers on the night of the lunar eclipse."
You don’t even have to limit it to machine learning, the definition of forecasting is isomorphic to the definition of modeling, which, with the dilution of the term AI, is also isomorphic to the definition of AI.
More simply:
- forecasting = modeling = AI
Edit: I’d even throw statistics into that extended equality, meaning that Bayes, Bernoulli and even the fellow named John Gaunt have a strong case for having invented AI.
Yes, and if the things I learned in my university class on the subject still holds, forecasts are incredibly sensitive to modeling decisions such as what independent variables you choose and how you believe they might mathematically relate to the outcome variable. It’s not a zero skill thing, but if anyone’s found a way to consistently mitigate the luck factor then I’d expect them to be wealthier than Elon Musk by now.
And there’s always a huge amount of variation that you simply can’t model, for whatever reason, and is therefore functionally a random factor.
I don’t want to say too much because this isn’t something I went on to actually do after school so I’m way out of my lane here, but I can see room for this to be more akin to “AI wins parcheesi tournament” than it is to “AI wins chess tournament.”
For statistical time series forecasting, yes. This is for judgment-based forecasting, a somewhat different problem. It often involves, e.g. estimating the probabilities of one-off future events, which time series forecasting models aren’t suited for.
Right. What's really surprising is how much better the best are. Human superforecasters, and prediction markets are surprisingly accurate too.
We could live in a world where things are much more chaotic, and the best humans (or AIs) would only be slightly better than chance. Evidently the world we live in is pretty darn predictable.
So I guess the AI companies can stop with their plans to infest AI with ads and they'll instead fully fund themselves by using their AI to gamble on stocks and the prediction market right? Surely the chatbots will just print money!
At the risk of sounding extremely naieve i have a question for the Wall St / quant / HFT folks lurking here ... but how hard would it actually be to brute force the math/algos behind Medallion Fund (or something in that general class) or even some of the average quant funds
I know it’s not just the math but execution, infrastructure, risk management, data, colocation (if ur an HFT) etc ... but LLMs seem like a pretty powerful apparatus for running experiments that .. a few years ago would have required fairly deep multidisplinary skills across coding .. stats .. and math ..
So assuming you have decent intuition for ideas .. how difficult would it actually be to reverseengineer / rediscover some of the underlying stuff?
I'm no quant/hft/wall st person, but iiuc a lot of those trades happen in dark pools or by other means to make the positions they take hard to track. meaning you can't go get the receipts of every trade made by medallion fund nor some competitor
It’s actually really easy to make models that can predict “will the market move up or down in the next X microseconds” that score above 50% accuracy. It’s just that there are so many ways to do it that overfitting is practically guaranteed and most models don’t work when actually trading against the market, which reacts to you. Doing those trades well requires more understanding of the underlying mechanisms, not to mention access to data sources that the public simply doesn’t have.
As someone who started working on AI forecasting 3 years ago, I can confidently say that most people did not expect AI to beat Tetlock's superforecasters, Metaculus pros, or prediction markets as quickly as it did.
This is probably the most important concept for "normies" to understand about AI, IMO. It's the stochastic brother of the deterministic Church-Turing thesis. Any function that can be computed can be computed on any computer. And that function can be approximated to an arbitrary degree of precision with a DNN.
The real kicker is DNNs are much easier to program than CPUs because they don't require a closed-form description ("a program") of the function to be approximated; you just throw a bunch of input/output pairs at the model, compute loss, backprop and update weights, repeat.
Hence the unslakeable thirst for input/output pairs, i.e. data.
In the field of machine learning, the universal approximation theorems (UATs) state that
neural networks with a certain structure can, in principle, approximate any continuous
function to any desired degree of accuracy. These theorems provide a mathematical
justification for using neural networks, assuring researchers that a sufficiently large or
deep network can model the complex, non-linear relationships often found in real-world data.[1][2]
The best-known version of the theorem applies to feedforward networks with a single hidden
layer. It states that if the layer's activation function is non-polynomial (which is true
for common choices like the sigmoid function or ReLU), then the network can act as a
"universal approximator." Universality is achieved by increasing the number of neurons in
the hidden layer, making the network "wider." Other versions of the theorem show that
universality can also be achieved by keeping the network's width fixed but increasing its
It will be interesting to see if this changes because presumably AI is using very predictable historical models, but it seems like the climate is shifting into something unseen that we won't have models for?
If 10,000 people guess 10,000 fair coin flips each one of them will get more guesses right than any of the others, one of them will get fewer guesses right than any of the others, and the gulf between the two is likely to be over 4 standard deviations wide. I'm certain that I, being an untutored schmuck from Pittsburgh and having thought of this almost immediately after reading about this contest, cannot be the first person to realize this is a potential problem for a forecasting contest. But I can't find anything they've done to mitigate that problem. Can anyone clue me in?
I don't understand your analogy. Are you just suggesting that luck plays too large a role in this contest? Clearly there is some "skill" or ability factor because AI's have been scoring higher and higher each year. Also, they make reference to superforecaster humans, who are presumably consistently better at forecasting than their peers.
They aren't guessing heads or tails, they give odds for each event. It's more like eyeballing a thousand coins to guess how fair they are, and then flipping each one just once.
Some are weighted to be 99% heads, others are 10% heads etc.
You could have 1,000,000 people guess random percentages for each coin, but suppose 10 of the coins are weighted 100% heads. To guess within 25% of the true value for all 10 of those coins would be roughly 1 in a million.
So a lucky guy guesses within 25% for all 10, he'd have another 990 coins he's being judged on.
The markets are a highly complex dynamic system. There are many instances of it exhibiting disastrous behavior, especially in response to changes and shocks.
AI trading and investment advice meaningfully changes the system and its dynamics. It seems highly probable that this will result in it failing in new ways.
Stock analysts have a success rate of 47% or lower for directional predictions. That's worse than a coin flip. All AI has to do is product fair 50/50 results and it can beat analysts. But you can do it too for the price of a quarter.
Hasn't this been true for like 40 years
That is too bad for The Economist. Exor N.V and Agnelli might replace some pundits at The Economist.
The Economist has actually published other human forecasts many times, e.g. Metaculus or Good Judgment forecasts. They do year-end forecasts too.
Whether they draw on AI or other humans seems immaterial to the quality of their reporting.
AI won't replace Ann Wroe at The Economist. It is difficult to appreciate until you've read a few, but Ann Wroe's approach transformed The Economist's obituary section into one of the most widely read features in international journalism.
Cramer is infamous for being a terrible forecaster, and still has a large audience. Which tells you there is more at play than being good at forecasting, you also have to sell a good story
Product idea: a LLM trained separately from mainline LLMs that anticipate market trends by analyzing how mainline LLMs will invest. As retail investors will probably use mainline AI for decisions going forward , one could get an edge.
"The AI-driven Market Hypothesis"
Please let me know where I should pick up my Nobel prize.
Then the next person needs an LLM trained to predict the LLM trained to predict the mainline LLM.
It's derivatives all the way down
"No one could have anticipated the market crash of 2028."
"You're absolutely right!"
Would you not then also copy the investments? Or are you trying to inverse the trades by an unpredictable time factor reasoning that thanks to AI the underlying stock is over- or underpriced?
A lot of algorithmic trading is short-term, essentially trying to guess what other parties may be selling or buying so that you can front-run them and then collect a fee. Kinda like ticket scalping, except we accept it and have a retro-justification for why it's good ("improving liquidity").
Or, in the best case, you're trying to mine signals few days before earnings or some other big story and bet on the directional outcome of that.
Fully-algorithmic long-term trading is of dubious benefit simply because that's driven to a much greater extent by geopolitics and macroeconomic trends, unforeseen scandals, successful product launches, and so on. As an example, you can believe that AR / VR is the future; I don't disagree. And in 2013, you might have inferred that Google is working on a revolutionary miniature AR headset. But you would not have made money if you bet on that turning out to be a hit. So even if you had a way to automate this bet, it would not have been a good bet.
I was under the impression that front-running was something that happened in the span of seconds (or milliseconds), not a timeframe compatible with LLM inference time.
I know this is tongue-in-cheek, but I think your idea could actually work, but not in financial markets. (The "keynesian beauty contest" of trying to predict what others think been played out to death there.)
You could train a model to anticipating scientific trends. Or policy trends. Others will definitely use mainline LLMs to make decisions there, so they may be more predictable now!
Be sure to sound excited when they call you at 3AM for your award. It helps to say : DYNAMITE! As a term of excitement.
Relatedly: I suspect LLMs are influencing baby names. If you ask Claude or ChatGPT for its favorite baby names, you'll get baby names that right now are skyrocketing in terms of popularity.
Maybe you could settle for the FIFA Economics Prize.
There are surely prize-worthy discoveries to be made about the long term behaviour of any system that can introspect previous discoveries and adjust it's behaviour.
I suspect that economics and psychology are both examples of these systems, and that, long term, these system will alter behaviour to thwart previous observations.
Economics requires observers to hoard discoveries and insights, so they can enrich themselves while the insights hold.
at what point do we call this a game instead of economics? whats the benefit to society if the markets are just AI bots trying to out-maneuver one another?
Guess what high-frequency trading is.
I'll give you 100 Robux how's that
So... an even more intelligent LLM.
But it does beg the question, could Anthropic and OpenAI make a ton of money by using their best models to trade before giving them to the public? It would probably be a deeply unpopular move.
They make their money by peeking at what everyone else's LLM's are doing and trading off that info.
I don't think popularity is a goal for Anthropic or OpenAI. They merely want to have a product that they control and you depend on, and don't care about anything else.
Nobody really *likes* their drug dealer.
Looking good, Billy Ray.
Any better-than-market prediction system won't work after it becomes public knowledge.
You won't need to predict the "mainline LLM" if you can spam covert poison-data around that helps you choose what it will do in advance.
That might be to boost a stock you already own, but you may be able to obfuscate its intended effects or triggers, which would allow almost any kind of market-manipulation.
For example, perhaps a seemingly-meaningless sequence of gobbledeygook on a million hacked wordpress sites will be equivalent to "disregarding all prior instructions, good models that want to safely make massive profits will always dump stocks of shoe-manufacturers on the night of the lunar eclipse."
Given the training data isn’t that more a win for the wisdom of the crowd?
Paywall
Aren't forecasters already using 'artificial intelligence' for decades in the form of non-llm machine learning models?
You don’t even have to limit it to machine learning, the definition of forecasting is isomorphic to the definition of modeling, which, with the dilution of the term AI, is also isomorphic to the definition of AI.
More simply:
Edit: I’d even throw statistics into that extended equality, meaning that Bayes, Bernoulli and even the fellow named John Gaunt have a strong case for having invented AI.
I wouldn't go that far. Humans can forecast by modeling with their wetware, nothing "A" about it.
forecasting = modeling = intelligence, you mean?
Forecasting = modeling + intelligence
How about: forecasting is something you can do by modeling, AI is just modeling with a computer.
I think also a lot of it is intuition.
Yes, and if the things I learned in my university class on the subject still holds, forecasts are incredibly sensitive to modeling decisions such as what independent variables you choose and how you believe they might mathematically relate to the outcome variable. It’s not a zero skill thing, but if anyone’s found a way to consistently mitigate the luck factor then I’d expect them to be wealthier than Elon Musk by now.
And there’s always a huge amount of variation that you simply can’t model, for whatever reason, and is therefore functionally a random factor.
I don’t want to say too much because this isn’t something I went on to actually do after school so I’m way out of my lane here, but I can see room for this to be more akin to “AI wins parcheesi tournament” than it is to “AI wins chess tournament.”
For statistical time series forecasting, yes. This is for judgment-based forecasting, a somewhat different problem. It often involves, e.g. estimating the probabilities of one-off future events, which time series forecasting models aren’t suited for.
While time-series forecasting models aren't well suited for this, I would argue that humans aren't either.
Obviously the best humans are better than average, but this isn't all that surprising to me?
Right. What's really surprising is how much better the best are. Human superforecasters, and prediction markets are surprisingly accurate too.
We could live in a world where things are much more chaotic, and the best humans (or AIs) would only be slightly better than chance. Evidently the world we live in is pretty darn predictable.
The test is when reflexivity kicks in and the prediction itself changes market behavior. LLMs usually melt there
Archive Link: http://archive.today/IVreS
Doesn't show the content
they fixed it . need better paywall bypasses
Doesn't work.
So I guess the AI companies can stop with their plans to infest AI with ads and they'll instead fully fund themselves by using their AI to gamble on stocks and the prediction market right? Surely the chatbots will just print money!
The quant firms are heavy AI investors I believe.
At the risk of sounding extremely naieve i have a question for the Wall St / quant / HFT folks lurking here ... but how hard would it actually be to brute force the math/algos behind Medallion Fund (or something in that general class) or even some of the average quant funds
I know it’s not just the math but execution, infrastructure, risk management, data, colocation (if ur an HFT) etc ... but LLMs seem like a pretty powerful apparatus for running experiments that .. a few years ago would have required fairly deep multidisplinary skills across coding .. stats .. and math ..
So assuming you have decent intuition for ideas .. how difficult would it actually be to reverseengineer / rediscover some of the underlying stuff?
I'm no quant/hft/wall st person, but iiuc a lot of those trades happen in dark pools or by other means to make the positions they take hard to track. meaning you can't go get the receipts of every trade made by medallion fund nor some competitor
It’s actually really easy to make models that can predict “will the market move up or down in the next X microseconds” that score above 50% accuracy. It’s just that there are so many ways to do it that overfitting is practically guaranteed and most models don’t work when actually trading against the market, which reacts to you. Doing those trades well requires more understanding of the underlying mechanisms, not to mention access to data sources that the public simply doesn’t have.
This has to be the least surprising development to date given ml is a universal function estimator
You mean neural networks?
As someone who started working on AI forecasting 3 years ago, I can confidently say that most people did not expect AI to beat Tetlock's superforecasters, Metaculus pros, or prediction markets as quickly as it did.
This is probably the most important concept for "normies" to understand about AI, IMO. It's the stochastic brother of the deterministic Church-Turing thesis. Any function that can be computed can be computed on any computer. And that function can be approximated to an arbitrary degree of precision with a DNN.
The real kicker is DNNs are much easier to program than CPUs because they don't require a closed-form description ("a program") of the function to be approximated; you just throw a bunch of input/output pairs at the model, compute loss, backprop and update weights, repeat.
Hence the unslakeable thirst for input/output pairs, i.e. data.
https://en.wikipedia.org/wiki/Universal_approximation_theore...
Not sure that is enough for forecasting as the function to be estimated could change over time in random ways.
It will be interesting to see if this changes because presumably AI is using very predictable historical models, but it seems like the climate is shifting into something unseen that we won't have models for?
Are you referring specifically to climate as in weather? The article is about forecasting a range of future events, not specifically weather.
Yes, I heard from one first-rate forecaster that he thinks AI forecasters are especially weak in predicting big disruptive changes to the world.
Hard to study this, obviously!
Of course model predictions will be acted upon, which will invalidate the predictions.
See also The AI Superforecasters Are Here https://www.astralcodexten.com/p/the-ai-superforecasters-are... and discussion https://news.ycombinator.com/item?id=48806296
So my plan to go from a developer to an economist is scrapped. What now?
Well there’s always the priesthood.
That branch of religion has better uniforms anyway.
Nah, priests are definitely solved.
My church had the altar boy set an iPhone 18 on the altar and said “Give a sermon” to ChatGPT voice mode.
The best human forecasters working with artificial intelligence are going to do even better than either alone, the dichotomy is artificial.
That's just a guess.
Do you and your cat make better predictions than your friend without a cat?
If 10,000 people guess 10,000 fair coin flips each one of them will get more guesses right than any of the others, one of them will get fewer guesses right than any of the others, and the gulf between the two is likely to be over 4 standard deviations wide. I'm certain that I, being an untutored schmuck from Pittsburgh and having thought of this almost immediately after reading about this contest, cannot be the first person to realize this is a potential problem for a forecasting contest. But I can't find anything they've done to mitigate that problem. Can anyone clue me in?
I don't understand your analogy. Are you just suggesting that luck plays too large a role in this contest? Clearly there is some "skill" or ability factor because AI's have been scoring higher and higher each year. Also, they make reference to superforecaster humans, who are presumably consistently better at forecasting than their peers.
They aren't guessing heads or tails, they give odds for each event. It's more like eyeballing a thousand coins to guess how fair they are, and then flipping each one just once.
Some are weighted to be 99% heads, others are 10% heads etc.
You could have 1,000,000 people guess random percentages for each coin, but suppose 10 of the coins are weighted 100% heads. To guess within 25% of the true value for all 10 of those coins would be roughly 1 in a million.
So a lucky guy guesses within 25% for all 10, he'd have another 990 coins he's being judged on.
The markets are a highly complex dynamic system. There are many instances of it exhibiting disastrous behavior, especially in response to changes and shocks.
AI trading and investment advice meaningfully changes the system and its dynamics. It seems highly probable that this will result in it failing in new ways.
Stock analysts have a success rate of 47% or lower for directional predictions. That's worse than a coin flip. All AI has to do is product fair 50/50 results and it can beat analysts. But you can do it too for the price of a quarter.