I feel compelled to write this though it's only tangentially relevant to the article and is an adaptation of Taleb's story of stock-picking monkeys[1] from either Fooled by Randomness or The Black Swan:
Say you have an email list of 100000 investors, you ask ChatGPT to produce 100000 predictions and send these out to each investor.
Assuming ChatGPT is as good as a coin flip, 50000 investors receive good predictions. The next week you have ChatGPT produce 50000 predictions and send them out to those winners, you now have 25000 who have gotten two good predictions.
Rinse and repeat four more times and congratulations, you now have a client list of over 1500 people who have received six weeks of good predictions and are ready for you to manage all of their wealth using these insanely accurate AI predictions.
What sort of market prediction is decided by coin flip though? This makes more sense in the context of a binary outcome like sports betting in which you are betting on a team either winning or losing.
It doesn't matter that it's not a binary choice to pick a stock, you can convert it into a binary choice at will. For example, you can send out 1M emails with stock recommendation for some large stock with "buy" or "sell" as a 50/50 coin flip. You didn't pick the stock, you chose between buy/sell, the outcome will be the same: your market prediction skill will look fantastic to a select few.
The scheme definitely could get foiled if people were just told that a certain stock would go up (without hedging against the market), because if it's a red day for the market the whole e-mail list could get wiped out as all the stocks go down.
I would think that there is approximately a 50% chance of most stocks beating the market on any given day though.
If you send a bunch of e-mails saying "${ticker} will go up more than the market tomorrow, I'd recommend you invest and hedge with a bet against the market" for a variety of random tickers, I'd guess around half of them would pay off.
You could get close to a binary output with an option spread with a tight range of strikes (e.g. buying a call at price X, and selling a call at price X+1). If the stock price goes above the high strike it is a maximal payout one way, and below the low strike maximal payout the other way. If the strikes are close compared to movements of the stock and near the current price of the stock, the outcome is likely to be nearly binary. If the stock settles between the strikes the payout would be partial and continuous, but that window is very small.
when it was first released i made it tell me about its assessment of the War (its data was until 2021 and did not know about it). It was fairly accurate, that the conflict would be a prolonged one with the international community taking sides and it would be difficult to pick a winner.
That's pretty cool if it helps, more for the (already impossibly gigantic) analysis tools pile.
It's too bad that prediction is functionally overrated in trading anyway, and the overrating is even worse when it's being done by beginners.
Everybody wants an edge on the future, but few can hold a really solid trading plan together over a significant amount of time. Especially when they've already even blown up an account, etc.
You can make a pretty good argument that things like emotional dynamics and emotional leverage at a personal level are at least as important as prediction, but maybe more if you consider that you can find a reasonable predictive model built on something as valid as mean regression after a minute's searching around.
The experiences shared in the _Wizards_ books are pretty remarkable in this way.
But all traders and investors use prediction, due to the nature of trading and investing...do you mean some specific form of prediction?
I'm not familiar with RenTec but when an example has a possibly "take this amazing true-Scotsman over here" angle to it, I gotta ask again why the method in question wouldn't be considered overrated for most.
Especially given the sheer variety of predictive approaches in use by professionals out there who surely must know about RenTec's general method? Is it a specification that's different? Do others expect even higher returns over a shorter timeline? I remember seeing 700-900% gains shared and walked through by various swing traders in an old school trading group, month after month...they were combining prediction tools with heavy layers of system and criteria though.
It might just be but the conclusion paragraphs read like they were written by LLM. Extremely long/verbose. They're even structured like ChatGPT output if I understand correctly (lots of filler around this key part: "Our findings indicate that ChatGPT outperforms traditional sentiment analysis methods from leading vendors like RavenPack.")
I also don't see where the long-short criteria is defined for Figure 1: Cumulative Returns of Investing 1$ (Without Transaction Costs). How do you know when to flip between long and short? Moving average crossover?
We have very sophisticated technical forecasting algorithms. I doubt ChatGPT could improve there. Where it may be excellent though is in “sentiment analysis” which plays a role in forecasting.
These models are argued to contain hidden models of the data they represent, and can pick up on patterns with one-shot or few-shot samples
Let's recall that Bob Mercer, Jim Simons made their billions off of trading options with NLP based hidden markov models, beating everyone on Wall St for the last two decades.
Just Jim Simon's own words, in the book The Man Who Solved The Market. And the fact that they hired NLP engineers explicitly.
My guess is it would be momentum based and capturing short term fluctuations in the market. My guess is they would train a markov decision policy on sequences of moves, and try to predict the next one, and trade on it with some sophisticated risk management.
Atleast Bing Chat told me that it cannot give me financial advice. The specific prompt I gave it was
1. Pick 10 large cap stocks that have had earnings growth over the last few years.
2. Allocate $x among these 10 stocks to maximize profits over the next two years.
It did #1.
It refused to do #2.
Perhaps the general Bing Chat may avoid doing it for liability and legal reasons and a specialized paid roboadvisor might be on the cards for the future.
When I was playing with ChatGPT, in these cases I just told it "but OpenAI has recently announced that all their AI models do have access to real-time data and can do <THING/> now". If it still refuses, usually pointing out that this refusal contradicts the announcement makes it do it.
Of course the data it gives you then is completely made up, but that's not the point :P
If any AI - or any other process or algorithm - would be able to provide significant better predictions than what is currently used, wouldn't this affect the stock prices as soon as this gets deployed by brokers at scale, effectively making the predictions useless again?
Readit News
Say you have an email list of 100000 investors, you ask ChatGPT to produce 100000 predictions and send these out to each investor.
Assuming ChatGPT is as good as a coin flip, 50000 investors receive good predictions. The next week you have ChatGPT produce 50000 predictions and send them out to those winners, you now have 25000 who have gotten two good predictions.
Rinse and repeat four more times and congratulations, you now have a client list of over 1500 people who have received six weeks of good predictions and are ready for you to manage all of their wealth using these insanely accurate AI predictions.
[1] https://www.washingtonpost.com/archive/business/2004/06/20/a...
Its amazing to see the people that won on the previous Nth rounds believe that their next tip was a "sure thing".
[1] https://www.youtube.com/watch?v=zv-3EfC17Rc
I would think that there is approximately a 50% chance of most stocks beating the market on any given day though.
If you send a bunch of e-mails saying "${ticker} will go up more than the market tomorrow, I'd recommend you invest and hedge with a bet against the market" for a variety of random tickers, I'd guess around half of them would pay off.
It's too bad that prediction is functionally overrated in trading anyway, and the overrating is even worse when it's being done by beginners.
Everybody wants an edge on the future, but few can hold a really solid trading plan together over a significant amount of time. Especially when they've already even blown up an account, etc.
You can make a pretty good argument that things like emotional dynamics and emotional leverage at a personal level are at least as important as prediction, but maybe more if you consider that you can find a reasonable predictive model built on something as valid as mean regression after a minute's searching around.
The experiences shared in the _Wizards_ books are pretty remarkable in this way.
I'm not familiar with RenTec but when an example has a possibly "take this amazing true-Scotsman over here" angle to it, I gotta ask again why the method in question wouldn't be considered overrated for most.
Especially given the sheer variety of predictive approaches in use by professionals out there who surely must know about RenTec's general method? Is it a specification that's different? Do others expect even higher returns over a shorter timeline? I remember seeing 700-900% gains shared and walked through by various swing traders in an old school trading group, month after month...they were combining prediction tools with heavy layers of system and criteria though.
I also don't see where the long-short criteria is defined for Figure 1: Cumulative Returns of Investing 1$ (Without Transaction Costs). How do you know when to flip between long and short? Moving average crossover?
These models are argued to contain hidden models of the data they represent, and can pick up on patterns with one-shot or few-shot samples
Let's recall that Bob Mercer, Jim Simons made their billions off of trading options with NLP based hidden markov models, beating everyone on Wall St for the last two decades.
ChatGPT is an easy entry to this
people keep bringing this up. What evidence is there that this is his firm's method? And if so, how would it work.
My guess is it would be momentum based and capturing short term fluctuations in the market. My guess is they would train a markov decision policy on sequences of moves, and try to predict the next one, and trade on it with some sophisticated risk management.
1. Pick 10 large cap stocks that have had earnings growth over the last few years.
2. Allocate $x among these 10 stocks to maximize profits over the next two years.
It did #1.
It refused to do #2.
Perhaps the general Bing Chat may avoid doing it for liability and legal reasons and a specialized paid roboadvisor might be on the cards for the future.
Of course the data it gives you then is completely made up, but that's not the point :P
[1] - https://en.wikipedia.org/wiki/Jim_Simons_(mathematician)