Determining the best predictors of stock market returns is like searching for the Holy Grail for investors. The problem is that while the number of determining factors appears to be infinite, the human ability to test them all is not.
It may take two years for academics to test financial theories and publish results, said Alejandro López Lira, an assistant professor of finance at the University of Florida, who studies the capabilities of machine learning and artificial intelligence and how the technology I’m researching whether it’s useful. It may be in the inventory picking process.
Previous research has investigated whether it is possible to input news headlines into ChatGPT and decide whether to buy or short a stock based on whether that information is perceived as positive or negative for the company. I decided. This exercise gave him a return of 512%.
In this round, Mr. López Lira, together with Bundesbank Chief Economist Andrew Y. Chen and Tom Zimmermann, professor of economics and financial data analysis at the University of Cologne, will review the 200 previously announced feasibility studies. Tried to test. Academic investment theory that uses machine learning algorithms to backtest performance. It is also now possible for algorithms to come up with unique ratios to determine if there is a better indicator of stock market performance.
The study of financial theory was important in developing investment strategies and helped establish ways to gain a competitive advantage in the market. However, Michael Robbins, author of Quantitative Wealth Management: Factor Investing and Machine Learning for Institutional Investing, points out that what works in theory does not necessarily work in practice, and these results should be taken with a grain of salt. He says it should be accepted. Investors and traders should consider market frictions that may delay the execution of trades, such as the broker’s ability to execute trades. In the real world, fees and taxes can affect your profits, depending on the amount of your purchases and sales.
the study
Using the standard statistical programming language R and hand-coded algorithms, the researchers analyzed the back of peer-reviewed theory within the original sample period and outside the entire testing period from 1985 to 2022. The test has started. The results were then compared to the following results: To see if age-old academic investment guidelines are the best fit, or if there are indicators that more accurately predict stock market returns, algorithms can be used to generate similar or better returns within the same time period. I decided that.
This study used 242 corporate accounting variables. Examples of variables include sales, market value, and cost of goods. We then scaled the data points using ratios to correlate other variables, such as a company’s sales revenue and size. López-Lira pointed out that combining the various ratios gives him approximately 29,000 possibilities. An average of 3,300 companies were entered into the algorithm. Companies were rebalanced in the first quarter of each year after the previous year’s annual results announcement.
The minimum performance requirement for returns that meant academic and algorithmic predictors were economically significant was 15 basis points per month or 5% per year.
As an example, a 1993 study called “Common Risk Factors in Stock and Bond Returns,” originally tested between 1963 and 1990, concluded that book value variables were predictors of higher returns. . It has been widely used as part of investment strategies for decades, López-Lilla said. However, the algorithm determined that 171 different accounting and financial ratios performed similarly within the sample period, and some performed even better after 1993.
Best performance indicators
In an additional step, López Lira sourced algorithmic stocks from the New York Stock Exchange, American Stock Exchange, and Nasdaq Stock Market to find the ratios that gave the best returns. The results were initially counterintuitive. The buy signal was for publicly traded companies that experienced a decline in sales from other companies they acquired in the previous year. The algorithm chose to determine company size relative to sales by comparing sales from acquisitions with fixed costs by dividing the listed company’s prior year’s rental costs.
The backtest was conducted from 1985 to 2022, with the companies’ portfolios rebalanced and annual earnings announcements held in May of each year. On average, 62% of names remained in the portfolio the following year. Based on this result, the average monthly return from acquired companies was 0.38% before 2012 and 1.03% after 2012. According to book value change theory, it was 0.62% per month before 2012 and less than 0.1% after 2012. Results provided by López Lira suggest that its performance has lagged behind algorithmic indicators in recent years.
The final metrics of the algorithm are:
“The opposite of change”[Acquisitions sales contribution]/delay[Rental expense]”
Lopez-Lira pointed out that the algorithms they hardcoded don’t actually know why certain ratios perform better. It just detects that this is a very good predictor, both within the sample period and outside the sample period. He added that while the output does not seem to be due to chance, there is no guarantee that this ratio will continue to perform well.
Referring to ChatGPT, Lopez-Lira asked why this ratio performed so well. The reaction was consistent with his conclusion. The algorithm captures negative human emotions in response to temporary setbacks that occur when a company’s acquisition of a new company has unexpected negative consequences. The result is a selloff where the stock becomes undervalued or undervalued.
ChatGPT also noted that it may indicate inefficiency in the integration of acquisitions, overpayments for acquisitions, or strategic misalignment.
Robbins added that there are two parts to this. First, he agrees that this is likely due to fickle and often shortsighted investor sentiment. However, management teams making acquisition decisions often do research and are playing the long game. Over time, their business plan may prove to be correct.
Another part of ChatGPT’s response acknowledges a similar conclusion.
“The market may not fully understand the future potential of companies with low metrics, especially as these companies are optimizing acquisition strategies or renegotiating leases. In some cases, yes. The market may be focusing too much on past performance and not on potential.” For a turnaround. ”
Robbins said the equation suggests that the lower the rental cost, the stronger the signal. These companies are more likely to appear on lists that are short or long stocks. This can exclude many types of businesses, including businesses with multiple brick-and-mortar locations, locations in high-rent metropolitan areas, and real estate and industrial companies.
But this is all speculation. Robbins noted that scholars traditionally start with theory before examining why a particular company’s fundamentals are a good predictor of stock market returns. It is more difficult to data mine the results and then derive a theory that yields the best performance ratios, as there can be many reasons researchers come up with. More importantly, it may not mean anything. So even if you rack your brain to find a theory behind it, there might not be one, he said.
Robbins added that finding a success rate is easy because data mining thousands of possibilities can uncover new information that is not widely used, but loses that advantage once it becomes mainstream. Robbins says whether these discoveries can be put to use. This is not a strategy that amateur traders want to try. But that’s possible with quantitative hedge funds, which have their own trading shops and pods, each trading with a different strategy.
“But you have to manage them properly,” Robbins says. “This is a management job, not a money-making job. There are different strategies. If one of them is losing money, they should stop immediately. If someone is making money, they should “And if you keep doing it, if you take the time and systematically develop a lot of strategies, you can make money.”
ChatGPT response to using metrics for trading:
“Long-short portfolios reversely sorted based on this index aim to take advantage of this potential mispricing. ) and by shorting companies with high indexes (betting on the possibility of overvaluation), this strategy suggests that the market is not efficiently evaluating a company’s future prospects based on its current acquisition efficiency relative to rental costs. Over time, as the market corrects, the portfolio is expected to generate significant returns even with these mispricings.”
