ÈÕ±¾avÊÓƵ

Zhiguang Wang, a researcher from ÈÕ±¾avÊÓƵ's Ness School of Management and Economics, is investigating artificial intelligence's ability to predict stock market returns. 

Like other sectors of society, artificial intelligence is fundamentally changing how investors, traders and companies make decisions in financial markets. AI models have the ability to analyze massive amounts of data while reading company filings or news headlines almost instantaneously. This is allowing for faster, more automated trading that is making it difficult for human traders — not utilizing AI — to find an edge in the markets.

One of the more interesting characteristics of AI is that the technology is advancing at near lightning speed. AI models that were "cutting-edge" just a year or so ago may be elementary when compared to current advancements. Past research has found neural networks, a computational machine learning model system, are among the best AI models for predicting the stock market. But is there a new AI system that can be more effective?

Zhiguang Wang is ÈÕ±¾avÊÓƵ's DuBois Professor of Business Finance and Investments in the Ness School of Management and Economics. This fall, Wang published a study, titled," which investigated if a newer AI architecture — transformers — could better predict stock market returns when analyzing economic data between 1957 and 2021.

Wang's research found that transformers significantly outperform neural networks at one-month, three-month and one-year intervals when predicting stock market returns.

"These results suggest that transformer architectures better encode fundamental information," Wang explained.

Transformers, which underly large language models such as ChatGPT, are an improved type of neural network that can process large data sets, including text. What makes transformer models particularly powerful is their ability to understand the context and relationships between words in a sentence. This allows them to extract deep, fundamental structure in the data and uncover low-frequency patterns that simpler neural networks would not be able to identify. In terms of stock market predictions, transformer models can pick up on long-term patterns and seasonality, allowing it to accurately forecast stock returns and improve upon previous AI-driven models.

"The transformer-based model can incorporate such macroeconomic variables as inflation, volatility indexes in credit, and economic policy uncertainty," Wang said. "Given the similarity in input features and the omnipresence of seasonality and autocorrelation in financial time series, the model can also be applied to other developed and emerging stock markets, and even to corporate bond markets."

Wang's study was published in the academic journal Finance Research Letters. Funding for this research was provided by the DuBois Endowment.