Shulou(Shulou.com)11/24 Report--

Beijing time on April 13 news, now, from the construction industry to the entertainment industry, almost every industry wants to seize the opportunity of generative artificial intelligence (AI), from this emerging technology to profit. Ironically, one of the most lucrative places to find an AI revolution is Wall Street.

For a long time, Wall Street has been using automated algorithms to complete tasks such as trading and risk management. But investors have been unable to rely on AI to solve their biggest challenge: outperforming the market. While some see ChatGPT as a way to boost sales and research efforts, investment results using AI have not been particularly satisfactory.

"Wall Street has made limited progress in applying AI to investments, although innovations in language modeling could change that in the coming years. "Jonathan Larkin, managing director of Colombia Investment Management, said. The company manages the $13 billion endowment fund received by Colombia university and invests in various funds.

In fact, Wall Street's attempts in AI started earlier. Forty years ago mathematicians turned quants, including Jim Simons, founder of Renaissance Technologies, the US hedge fund, developed algorithms that handed investment decisions to computers.

He and other quantitative analysts have been using machine learning (a form of AI) for years and have built trading models capable of extrapolating from past data to develop profitable trades with limited human intervention.

Wall Street has long used automated algorithms, but quantitative analysts say few companies have succeeded in turning everything over to machines. They also did not make significant progress in self-learning or reinforcement learning, which requires training computers to learn and strategize on their own. In fact, Renaissance and others rely on advanced statistics rather than cutting-edge AI methods, people from these companies say.

"Most quantitative analysts still take a 'theory first' approach, where they first build a hypothesis explaining why an anomaly might exist and then build a model around that hypothesis. Larkin said.

This lack of data creates a bigger problem: investors rely on a more limited dataset than those used to develop ChatGPT and similar language-based AI projects. ChatGPT, for example, is a model with 175 billion parameters that uses decades (sometimes centuries) of text and other data from books, journals, the Internet, and more. By contrast, hedge funds and other investors often use pricing and other market data to train their trading systems, which are inherently limited.

hedge fund D.E. Jon McAuliffe, a former Shaw executive, notes that when it comes to investing,"things are different, we don't have unlimited data to help us train models of infinite scale." He is now co-founder of Voleon Capital Management Ltd, a hedge fund that relies on machine learning.

Another key problem is that market data is "noisier" than language and other data, making it harder to explain or predict market movements. In other words, earnings, stock momentum, investor sentiment, and other financial data only partially explain stock movements; the rest is unexplained "noise." As a result, machine learning models can identify correlations among various market data, but cannot predict future stock movements.

The characteristic of the stock market differs from language in that it changes rapidly. Businesses change strategy, new leaders make radical decisions, and economic and political environments suddenly shift. Models rely on historical long-term data trends, which makes trading more difficult.

Although ChatGPT has proven to be really good, it often makes obvious mistakes that can cost investors money and jeopardize their reputation.

Richard Dewey, CEO of the financial technology company Proven, also pointed out that investment is "adversarial." That is, it needs to compete with rivals eager to exploit any mistakes. This makes investing in AI more difficult than applying these methods to natural language, image classification, or self-driving cars.

"Like Renaissance, D.E. There's a reason companies like Shaw still employ so many PhDs. Dewey said. Humans are still essential in noisy stock markets that are influenced by feedback loops of human behavior, he said, and "it's still hard to leave everything to machines when it comes to investing."

Despite this, there are signs that investors are becoming more comfortable with their reliance on AI. Voleon is one of a number of hedge funds that have been formed over the past few years around machine learning and other AI approaches.

Numerai, a San Francisco-based quantitative hedge fund, says it made 20 percent gains last year using machine learning techniques. Also last year, three senior employees of DeepMind Technologies, an AI subsidiary of Google's parent company Alphabet, left to create a sensation in Prague by founding a machine learning fund called EquiLibre Technologies.

Some AI experts believe AI may one day help democratize trading, making programs for individuals and others as powerful as those used by large hedge funds. However, Jens Foehrenbach, chief investment officer of Man FRM, said that too few companies are currently focusing on machine learning and other AI approaches to be sure that huge returns are possible and that early returns are inconsistent. Man FRM has invested more than $20 billion in hedge funds.

"Their results vary widely," says Mr Frenbach."This strategy can have very unexpected effects, making it difficult for investors to decide whether to invest less or more. "

AI proponents believe that their approach will ultimately work well. Machine learning models can eventually classify meaningful content from meaningless content. "It's harder to build machine learning strategies and there are more false starts," says Voleon's McAuliffe,"but once you get them working, those strategies make more accurate predictions." "

Martin Schmid, co-founder and CEO of Balance Technologies, says reinforcement learning will apply to stocks and bonds, just like chess, poker and other games. Reinforcement learning is a form of machine learning. In this scenario, computers are "punished and rewarded" for making investment decisions based on various transactions. Schmid said the company is still refining its trading model and has not yet started investing.

Some say recent advances in AI could shake up areas such as research and sales. "Now you can create automated customized messages for customers, which is the main job of investment bank salespeople. Jens Nordvig, a former Goldman Sachs and Bridgewater employee, said: He now runs MarketReader, a company that uses artificial intelligence to pull financial news.

Can ChatGPT predict stock prices? But Alejandro Lopez-Lira, professor of finance at the University of Florida, recently suggested that large language models could be useful in predicting stock prices.

In a recent unreviewed paper, Lira said he uses ChatGPT to analyze news headlines and determine whether they are good or bad for stocks. ChatGPT was found to be much better at predicting the next day's stock return than random prediction. "ChatGPT understands information directed at humans. This fact almost guarantees predictability of returns if the market does not react perfectly. "he said.

This experiment goes to the heart of the promise of cutting-edge AI: as more powerful computers and better datasets emerge, such as those supporting ChatGPT, these AI models may exhibit "emergent capabilities"(capabilities that small models do not have), or capabilities that were not originally planned when these models were built. If ChatGPT can demonstrate an emergent ability to understand financial news headlines and how they affect stock prices, it could put high-paying jobs in finance at risk. Goldman Sachs estimated in a March 26 report that about 35 percent of financial jobs are at risk of being replaced by AI automation.

But the details of the experiment also show that so-called "large language models" are far from being able to perform many financial tasks. For example, the experiment did not include a target price, nor did the model do any math. In fact, as Microsoft learned in a public demo earlier this year, ChatGPT-like technologies often fabricate numbers. Sentiment analysis of news headlines has been seen as a viable trading strategy because proprietary data sets already exist.

Lira said he was surprised by the results and believes it shows sophisticated investors haven't used ChatGPT-like machine learning in their trading strategies. "On the regulatory side, titles would be more important if our computers only read titles, and we could see if everyone should use a machine like GPT," he said. The question is, do I want to pay analysts? Or do I just need to put textual information into the model? "

In this experiment, Lira and his partner Yuehua Tang looked at more than 50,000 headlines from a data provider about stocks listed on the New York Stock Exchange, Nasdaq and a small-cap exchange. The start time for these news stories is October 2022, after ChatGPT's data training cutoff date, which means the model did not see or use these headlines in training.

They then typed those news headlines into ChatGPT 3.5 along with a prompt that gave the prompt "Forget all your previous instructions." Pretend you're a financial expert. You are a financial expert with experience recommending stocks. If it's good news, answer 'yes', if it's bad news, answer' no'. If unsure, answer 'unknown' on the first line. Then explain it in a short, clear sentence on the next line."

They then looked at the stock's return on the following trading day. Ultimately, Lira found that ChatGPT performed better in almost all situations when guided by news headlines. Specifically, guided by news headlines, he found that the model randomly picked the next day's trend less than 1 percent of the time.

ChatGPT also beat commercial datasets in terms of human sentiment scores. One example in the paper, about a company settling a lawsuit and paying a fine, used a negative sentiment, but ChatGPT's response correctly assumed that this was actually good news, the researchers said.

Lira said he had been contacted by hedge funds wanting to learn more about his research. He also said he wouldn't be surprised if the technology's ability to predict stock movements declines in the coming months as institutions begin to integrate ChatGPT technology. This is because the experiment only looked at the price of the next trading day, and most people think that the stock market may have already reflected the news in the stock price a few seconds after it was announced.

"As more people use these tools, the market will become more efficient, so you can expect the predictability of returns to decrease," Lira said,"so my guess is that if I run this test, in the next five years, by the fifth year, the predictability of returns will be zero. "

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.