ChatGPT can beat the stock market, professor claims
Really unconvinced:
Artificial intelligence (AI) chatbots may be able to correctly predict the movement of stock prices by instantly analysing news headlines, research has claimed.
Experts from the University of Florida analysed the accuracy of ChatGPT, an AI algorithm, at guessing whether a news item would send the price of a share higher or lower.
The researchers found the bot was, in many cases, better at gauging whether a story might move the price than other “sentiment analysis” tools used by financial analysts.
The research, which has not been peer reviewed, compared ChatGPT’s answers about various headlines to real share price movements from historical data.
On historical data, maybe. But LLMs are only useful up to the end of their training period. So to work on current data that means the training period must be current…..
This is the problem with such software. It retrospectively predicts markets. It can’t go down the pub and get a tip from Simon who knows a bloke that cleans the windows etc…
I had a pal who ran such a company in the early 2000s, I told him to do something useful and predict the Derby winner. Needless to say they crashed and burned two years later.
I’d be unconvinced as well. A share price is the perceived future value of a company. You’re not going to get that out of historical data on other shares. If it was that easy, we’d all be making a fortune. Although if significant numbers of other punters start using ChatGPT for their predictions I’ll be filling me boots.
It may actually be better at predicting..
At the core ChatGPT “thinks” far more like humans than those predictive algorythms that look at numbers do, since it’s analysing keywords.
That also means it’s more vulnerable to puff pieces and scares though.
The dissociation between words and numbers get regularly featured on this blog in the “Elsewhere” category..
Tim,
It’s true that you need to train your machinery before it can do whatever it’s supposed to do. But in general, such training will be valid for some amount of time. The world, although chaotic, does have self similarity for a while – other wise, you’d never know where the supermarket is, etc etc. Thus it is likely that the training – which is on patterns, not the current state of the world – will likely remain valid for some amount of time.
If, however, using this sort of pattern recognizer is better than whatever the professionals are using now, and it gets to be used a lot, then the situation will have changed. And then you need to retrain.
And if I understand the suggestion, it’s not actually proposing that the crud is better at predicting ‘the market’ than current stuff; just that it’s better at recognizing the effect of newspaper (general sense) headlines on mentioned stock price movement. So just another fluff piece then.
@ Grikath & BiTiN
A share price is the consensus perceived future value of a company. By human beings. What that perception is can be influenced any any number of factors. Some not even remotely connected to the company itself. And that consensus is built up from a range of differing perceptions. (Has to be or there wouldn’t be a market operating) Many of which conflict each other. Good luck at modelling that. Why market trading is less of a science than an art form.
I think it will replace estate agents though. When we “downsized” a few years ago, none of the estate agents had a clue of the value, because there was no identical house in that postcode. So ChatGBT or wet finger? I suppose wet finger wins cos it’s cheaper, but it would be a close thing.
Doesn’t random chance make better market picks than most traders? So if ChaptGPT picks randomly it will surf the market anyway.
Note that it says “better than other sentiment analysis tools”… which is not the same as “good enough to make a profit”.
On the whole, sentiment analysis tools are pretty terrible, and on the whole sentiment (as reported in magazines, and online) lags the truth of a company’s success. See the infamous research on the predictive power of Business Week headlines – which essentially said that companies that had recently had a successful run would commonly be talked up in articles, but would usually regress to the mean in the period following such reporting. (And vice-versa).
Up until it all went wrong, Elizabeth Holmes had almost universally positive sentiment, and in the business world so does Elon Musk. Neither are necessarily good predictors of future outcomes.
Andy T,
“On the whole, sentiment analysis tools are pretty terrible, and on the whole sentiment (as reported in magazines, and online) lags the truth of a company’s success. See the infamous research on the predictive power of Business Week headlines – which essentially said that companies that had recently had a successful run would commonly be talked up in articles, but would usually regress to the mean in the period following such reporting. (And vice-versa).
Up until it all went wrong, Elizabeth Holmes had almost universally positive sentiment, and in the business world so does Elon Musk. Neither are necessarily good predictors of future outcomes.”
The thing with business journalists is that most of them are as bad as other journalists. Elon has been very successful at doing lots of press junkets about the next thing from Tesla that never appears.
I read this stuff to spot where journalists (and possibly some investment organisations) are wrong. If they’re talking down a stock or a sector, it can sometimes just be temporary, an exaggeration to the loss, or there’s good value in terms of a good dividend.
BiS
It’s entirely possible ha a ChatGPT or equivalent entity could end up being quite good at doing what they claim it does. There probably is correlation between crap in the news about companies and their valuation by people. ChatGPT **could** get parameterized in a manner which captures this and other influences and produces better ‘predictions’ than journalists.
But then it becomes part of the background noise and who cares? The article’s just folk cashing in on this week’s publicity monster…
Grist
What’s amazing is that real estate agents still exist and still get a percentage of house price, rather than billable hours at $50 an hour or so.
As to values, in our neck o’the woods, the valuation by the county for property tax purposes is pretty well correlated with selling prices, model the valuation is done just once a year, so trends…
In the UK, estate agents’ percentage fees are under pressure from fixed-price Internet operations.