Machine learning forecasts of corporate earnings outperform analyst forecasts, by revealing new information, economically important predictors and capturing non-linear relationships. Investors can use ML models as a less-biased forecast and a decision-making tool for when there is a vacuum of analyst coverage, an award-winning research paper has found.
Machine learning (ML) forecasts of corporate earnings have been proven to be significantly more accurate than analyst coverage over long horizons, thanks to the technology’s ability to capture subtle, non-linear interactions between economic data points, new research has found.
The paper, Fundamental Analysis via Machine Learning, which won the 2024 Graham and Dodd top award from CFA Institute, highlights ML model’s potential for use by investors and “considerable promise” to save costs and enhance efficacy in fundamental analysis.
Co-authored by independent researcher Kai Cao and Haifeng You, Chair Professor of Accounting at Shenzhen International Graduate School at Tsinghua University, the research combined three popular algorithms including decision trees and artificial neural networks to create an ML forecasting model, and trained it on a comprehensive set of financial statement items.
The model is then used to generate forecasts for firm data – outside of those used in its training – from 1975 to 2019. The paper found that the ML model performed well against analyst forecasts produced over similar time periods, even though the latter usually has access to more information than just financial statements.
“ML forecasts are as accurate as consensus analyst forecasts over a one-year forecast horizon and are significantly more accurate than them over longer forecast horizons,” the study said.
“And ML forecasts contain significant incremental information beyond analyst consensus forecasts even if analysts have access to all the financial statements used in ML models (and much more), suggesting that analysts fail to fully incorporate the information in key financial statement items into their forecasts.”
The ML model in the research was trained with 60 data points, which include companies’ historical earnings; advertising and R&D expenses; individual balance sheet items such as assets and liabilities; and operating cash flows.
“Corporate earnings are the cumulative result of a myriad of transactions, each reflected within various financial statement items that can have disparate impacts on future earnings,” the research said. But ML’s ability to capture subtle, non-linear interactions between economic data points may have contributed to better forecast results.
This presents not only an opportunity for investors to use ML models as an alternative and “less biased” forecast compared to that of analysts and stock pickers, but also as a decision-making tool when there is a vacuum of analyst coverage – for when a company only operated for a short period, for example.
Other parts of the research highlighted the ML model’s superior accuracy when compared against six other existing earnings forecast models such as a random walk model.
“Cross-sectional analyses indicates that the ML model leads to even greater accuracy improvements among firms with more difficult-to-forecast earnings,” the research said.
“We then test whether the new information uncovered by the ML model can lead to significant improvements in investment decision-making… The results show that the new information component has significant predictive power with respect to future stock returns.”
The research noted there is room for an even more powerful application of the ML model if sophisticated investors integrate it well with their risk and transaction models and other portfolio optimisers.
Meanwhile, Empirical Evidence on the Stock–Bond Correlation, co-authored by Edouard Senechal, senior portfolio manager at State of Wisconsin Investment Board won the 2024 Graham and Dodd Scroll Award.
Presenting his findings at the Top1000funds.com Fiduciary Investors Symposium at Oxford last November, Senechal identified macro variables such as inflation and real rates have a material impact on stock-bond correlations. He warned investors not to take past correlations between asset classes as a given in the future.
“Most people use data from the last 20-30 years, but that is not necessarily reflective of what we will get in the next 20-30 years,” Senechal said.
