Chicago Booth Review Podcast Three Ways A.I. Can Improve Decision-Making

Hal Weitzman: In the perfect world of economic models, investors make perfectly rational decisions using perfect information and earn the best possible returns. They never get distracted or confused.

Of course, in real life it doesn’t always work that way. Investors can get bamboozled by management drivel, or besotted with charismatic founders.

What if we could use AI to make better investment decisions? Welcome to the Chicago Booth Review Podcast, where we bring you groundbreaking academic research in a clear and straightforward way. I’m Hal Weitzman, and in this episode we explore how researchers are using machine-learning models to improve how investors allocate their funds.

Method #1 involves using Chat GPT to help investors make better decisions by cutting the corporate business speak out of company filings. Chicago Booth researchers applied the chatbot to summarize wordy and complex regulatory filings, and found that by getting to the point, it could enable investors to make snappier and more accurate decisions. The piece was written by regular CBR contributor Rebecca Stropoli, and it’s read for us by Julie Granata-Hunicutt.

Reader: If you read the headlines, large language models such as ChatGPT could be set to transform industries, eliminate millions of jobs, or even destroy humanity as we know it. How much of this is hype remains to be seen, but research from Chicago Booth PhD student Alex G. Kim and Booth’s Maximilian Muhn and Valeri Nikolaev finds that ChatGPT, a chatbot with conversational applications, could be good for investors, who stand to benefit from its ability to cut the fat from corporate disclosures.

The researchers focused on two key sources for investors researching a company’s stock: MD&A disclosures—in written annual or quarterly reports, the section that provides management discussion and analysis of a company’s performance—and earnings conference calls, which dig into reported results and allow analysts to ask questions of executives.

Regulators have expressed concern that company disclosures contain information that’s too complex for many investors to comprehend, and that companies might hide negative news with redundant and irrelevant language.

The researchers put the ChatGPT model to the test to see if it could slash the length of corporate disclosures and provide summaries that were both fat free and effective. They gathered the MD&As and conference-call transcripts of all public US companies from 2009 to 2020, and then took a sample that included 1,790 MD&As from nearly 340 companies and about 8,900 conference calls from almost 370 companies. The researchers then prompted the 3.5 Turbo version of ChatGPT to summarize all the reports without referencing any outside information.

The summarized MD&As and conference-call transcripts were, on average, substantially shorter than the original documents—less than half the size. Cutting out cautious language and generic statements contained in many of the longer reports provided more accurate assessment of the genuine sentiment—whether positive or negative—expressed in corporate disclosures, the research finds.

Kim, Muhn, and Nikolaev then studied the companies’ individual stock activity over a two-day window following their disclosures. They find that the clearer sentiment of the summarized reports better explained the price movements that followed disclosures. Additionally, when a company reported losses and displayed negative sentiment in its original report, the amount of bloat (that is, the redundant, irrelevant language not included in the summaries) was generally higher—consistent with the idea that companies seek to obscure downbeat news.

More bloated reports were also associated with a drop in overall price efficiency for a company’s stock. Prices are most efficient when both parties in a stock trade have equal information—and less efficient when one party is more informed than the other. A professional manager might use tools to analyze long reports, but a typical retail investor doesn’t have access to those. Through an estimate based on the statistical model, the researchers find that a 1 standard deviation increase in bloat for MD&As led to a 1.6 percent increase in the probability of “informed trading” (one party being more informed that the other), and thus less-efficient trading.

They also find the added bloat led to a 9 percentage point decrease in the speed of price discovery for a company’s stock, and an 18 percentage point increase in the bid-ask spread (the difference between the maximum price buyers are willing to pay and the minimum price sellers are willing to accept). The higher the bid-ask spread, the more expensive it is to trade.

For conference calls, the analysis indicates that a 1 standard deviation increase in bloat led to a 1.7 percent increase in the probability of informed trading, a 5 percentage point decrease in the speed of price discovery, and a 45 percentage point increase in the bid-ask spread.

Finally, the researchers find that ChatGPT can help investors target specific areas of interest within long financial reports. For example, if they are interested in understanding a company’s dedication to environmental, social, and governance policies, they can prompt the chatbot to target that information and get a clearer view.

Both institutional and retail investors could benefit from ChatGPT’s ability to get to the point of the reports, says Kim, who adds that he is not sure yet whether ChatGPT in its current form can directly help investors build better portfolios or uncover the most lucrative stocks. Instead, he says, it could save them the time and effort of poring over complex documents and help them to better process the most important information. Thus, when making a trade, they can at least have a clearer idea of the fundamentals behind it.

Hal Weitzman: Method #2 is another application that could help technical stock-market investors, especially those working with incomplete data sets. Researchers at the University of Chicago and Yale are applying the machine-learning programs that are used to help self-driving cars identify pedestrians and stop signs to spot patterns in stock-market data, even if the datasets are partial or very short.

Regular CBR contributor Michael Maiello dug into the research in a 2022 piece entitled, “How Self-Driving-Car Technology Can Help Machines Trade Stocks.” It’s read for us by Julie Granata-Hunicutt.

Reader: Traders have long used price charts to predict future investment returns. Patterns such as the “head and shoulders,” “double top,” or “ascending triangle” form a visual shorthand for investors looking for repeatable shapes and figures that suggest profitable trades. Now artificial intelligence, using the same pattern-recognition techniques that factor into programming self-driving cars, is proving adept at technical investing as well.

University of Chicago PhD student Jingwen Jiang, Yale’s Bryan Kelly, and Chicago Booth’s Dacheng Xiu built a machine learning–based program that can recognize the patterns in stock-price charts and use them to generate profitable trading strategies. The findings support the idea that visual representations of stock prices contain valuable and actionable information.

“We have heard that traders can translate price charts into signals,” Xiu says. “Our findings suggest that the shape of the data contains enough information for trading. This justifies what technical traders do.”

Jiang, Kelly, and Xiu find that using past prices and momentum, expressed graphically, can generate profitable trading strategies over short time periods. The effect wanes the longer a position is held. As the holding periods grow to months and quarters, fundamental issues such as financial reporting and corporate events begin to hold more sway over future prices. They find that their chart-reading ML system’s predictions delivered higher returns with less risk than trades made using other methods for predicting returns.

The A.I. methodology the researchers employ is called a convolutional neural network (CNN). It’s the same type of ML that programm¬ers are using to help autonomous vehicles identify pedestrians, stop signs, and sidewalk curbs. Its performance in recognizing investment patterns suggests that computers can reliably predict returns without relying on the intuitions, hunches, and judgments that human investors develop over years of trading.

“A technical trader uses prior knowledge to define patterns,” Xiu says. “A CNN has no prior knowledge. Without using any existing chart, I am going to ask the CNN to learn from the price curve to extract useful information for prediction.”

The findings also suggest that a CNN can make useful predictions even when data sets are incomplete or small. The researchers find that they can use patterns observed in data-rich US-based markets to make predictions about stock-price performance in foreign markets with shorter track records.

Suppose you have two data sets, Xiu says. One is high quality and can be used to train a computer to recognize a cat. The other, more limited one contains data about tigers. Essentially, the researchers find that they can use lower-level features associated with both cats and tigers, primarily learned from the first data set, to teach a computer to recognize a tiger, with little or even no information from the second data set.

Through this transfer learning, the researchers say, a CNN system might be able to analyze newer and emerging asset classes if they share features with more mature markets.

Hal Weitzman: Method #3 is about using A.I. to help improve decision-making by investors in startups. Most venture capitalists tend to invest in lots of new companies, knowing that the vast majority will fail to make any real money. But research suggests that using AI could improve the proportion of winners, by focusing on numbers and shutting out the distraction of personalities. CBR contributor Sarah Kuta explained the research in a 2022 piece entitled, “How Startup Investors Could Back More Winners.” It’s read for us by Julie Granata-Hunicutt.

Reader: On TV shows such as Shark Tank and in private boardrooms across the world, startup founders passionately make the case for why their fledgling companies are the next Ubers or WhatsApps. While the business cases they make are critical, so is the ability to pitch with charisma, clarity, and a personal story.

But venture capitalists tend to rely too heavily on founders’ backgrounds when investing in startups, suggests Princeton postdoctoral researcher Diag Davenport (a recent graduate of Chicago Booth’s PhD Program). And they could make more money—10 percent more, according to Davenport—by relying less on their impressions and more on data to make investment decisions, he argues.

Comparing real investor choices with an algorithm’s predictions, Davenport finds that up to half of VC investments were “predictably bad,” he writes.

“A purely intuitive, qualitative approach to investing is more costly than they likely realize,” Davenport says. “Relying on data is a way to make their decisions more objective and less vulnerable to mistakes.”

Davenport identified every company that participated in any of the top 100 startup incubator and accelerator programs between 2009 and 2016. Some eventually became household names—Instacart and Uber, for instance—whereas others never went anywhere. He created a data set of all the equity deals venture capitalists made backing the 16,054 companies within the first five years of their completion of an incubator or accelerator program. Then he tracked their valuations over time as well as whether they made a successful exit via an initial public offering, a merger or acquisition, or any Series C or later funding round.

He trained an algorithm to make predictions about the companies by analyzing a wealth of data, including the financial information available at the time of the incubator or accelerator program, the founders’ backgrounds, the companies’ missions, and the ways they made money.

When Davenport contrasted the algorithm’s predictions with the investors’ decisions, he finds that the venture capitalists spent millions of dollars to back startups they should’ve known would inevitably fail. For his data set covering $9 billion of investments, the lost returns came to $900 million, he calculates, reaching that conclusion after modeling alternate portfolios that dropped low-performing startups and replaced them with bonds or stocks.

There was a strong link between decisions to invest and startup founders’ backgrounds—details such as where the founders went to school or how many board seats they occupied, the study indicates. This was especially the case when investors were considering low-quality startups. A founder’s background and experience do matter, but “the algorithm tells us venture capitalists rely on these things too much,” Davenport says.

The good news is that they could easily improve their decisions by incorporating more machine-learning techniques, which would benefit not only themselves but their limited partners (including their clients, university endowments, and retirement pensions), who rely on them to make savvy financial moves.

At least for now, however, most venture capitalists are reluctant to use artificial-intelligence tools. For one thing, they worry about the reliability of algorithms, which use past patterns to make projections. Data on startups are also notoriously difficult to obtain. But perhaps the biggest reason is pride.

“They’ll say, ‘Look, you have to be in the room and see the passion in the eyes of the founder—no algorithm could ever detect that,’” Davenport says. “That highlights the overconfidence a lot of investors have, particularly because they already have good returns. But they underperform what they could be getting. It creates this situation where, if you win the game, it’s hard to understand how many more points you could’ve scored.”

Hal Weitzman: There’s a lot more about A.I. and its applications on CBR’s website at chicagobooth.edu/review. That’s it for this episode of the Chicago Booth Review Podcast. It was produced by Josh Stunkel. If you enjoyed this episode, please subscribe and please do leave us a 5-star review. Thanks. Until next time, I’m Hal Weitzman. Goodbye.