An earnings forecast that proves to be even moderately off target on the low side can damage a company’s stock price in the short term, and the effects of big misses can last for weeks or months.
Unfortunately, misses are common, and investors tend to punish them much more than they reward profits that beat forecasts. That prospect may be on CFOs’ minds even more than usual lately, as shares of companies that underperformed earnings expectations in 2025 fared particularly poorly compared to historical averages.
One path to improved forecasting accuracy that’s gaining favor entails using artificial intelligence or machine learning models to track the relationships between known profitability drivers and actual results.
Just how much of a difference can that make? A just-published academic paper details a forecasting methodology incorporating machine learning that reduces mean absolute average forecast errors by approximately 7%, compared with the commonly used “random walk” forecasting method.
Oliver Binz, an assistant accounting professor at German business school ESMT Berlin who co-authored the paper, noted that traditional forecasting models, including random walk, assume simple, straight-line relationships between profit-driving variables and actual profitability that do not reflect the non-linear manner in which they interact in the real world.
“Researchers have struggled to develop models that do better than simply assuming that ‘next year will look like this year,’ as random walk implies,” said Binz. “This paper shows that financial statements contain more predictive information than we thought, but only if we analyze them using tools that match their complexity.”
According to Binz, models not utilizing AI/machine learning have generally taken a “kitchen-sink” approach with no systematic selection of the variables used as predictors, meaning they include many more variables than those that are useful for predicting profitability.
The “breakthrough” accomplished by the research, said Binz, was enabled by combining a proven structured accounting framework for profitability decomposition, together with a modern machine learning algorithm for predictive analytics.
Instead of “throwing thousands of variables into an AI model,” the profitability analysis — developed by Nissan and Penman and based on the classic DuPont financial analysis model — breaks profitability into clear building blocks, including operating performance, financial leverage, profit margins, asset efficiency and core vs. one-off items.
For the predictive piece, the researchers used gradient-boosting regression trees, a method that has done well in other research efforts around the use of AI/machine learning for financial forecasting, according to Binz.
The researchers trained the model by applying it to all financial statements of publicly traded U.S. companies from 1963 through 2023, using return on common equity as the measure of profitability.
Using either the AI/machine learning algorithm or the structured accounting framework alone, without the other, yielded much less accurate forecasts, Binz noted.
The broader message for CFOs is not that artificial intelligence replaces financial expertise, but that it finally allows that expertise to scale.
“As forecasting becomes both more accurate and more explainable, finance leaders gain a powerful tool to navigate uncertainty, manage investor expectations and lead with greater confidence in volatile markets,” said Binz.
The paper, co-authored by Binz, Duke University’s Katherine Schipper and the University of Utah’s Kevin Standridge, was published in the December 2025 issue of the Journal of Accounting and Economics.
