Artificial Intelligence and Machine Learning for Stock Market Prediction: A Systematic Literature Review of Methods, Emerging Paradigms, Reproducibility, and Evaluation Gaps (2016-2026)
This systematic literature review of 685 studies (2016–2026) analyzes the evolution, reproducibility, and evaluation gaps in AI-driven stock market prediction, revealing a critical disconnect between reported model accuracies and practical trading utility due to widespread issues like data leakage, lack of code sharing, and insufficient statistical testing.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Stock prices are notoriously difficult to predict. They do not move in straight lines, they are often noisy, and they change their behavior over time. For decades, economists tried to forecast these movements using simple mathematical rules that assumed the future would look much like the past. But as computers became faster and more powerful, a new generation of researchers began using artificial intelligence to find patterns in the chaos. These systems, known as machine learning models, can digest vast amounts of historical data to spot subtle signals that human analysts might miss. The promise is tantalizing: if a computer can learn to predict which way a stock will move, it could generate significant wealth. Yet, despite the excitement, the field has been plagued by a persistent problem. Many studies claim to have found a winning formula, but when others try to repeat the work, the results often vanish.
A team of researchers set out to understand why this gap exists between what papers report and what actually works in the real world. They conducted a massive review of 685 scientific studies published between 2016 and 2026, examining how artificial intelligence is currently being used to forecast stock markets. Their work is not a new prediction model itself, but rather a deep audit of the entire field. They looked at the types of computer architectures being built, how the methods have changed over the last decade, and, most critically, whether the results are trustworthy. By sorting through thousands of pages of research, they uncovered a landscape where the technology is advancing rapidly, but the standards for proving that technology works have not kept pace.
The researchers found that the field has moved far beyond simple linear equations. In the early years of their review, studies relied on older statistical tools or basic machine learning techniques. By the 2020s, the dominant approach shifted toward deep learning, specifically a type of neural network called a Long Short-Term Memory, or LSTM. These networks are designed to remember sequences of events, making them well-suited for time-based data like stock prices. In fact, nearly 90 percent of the studies they examined used some form of this technology. More recently, the field has begun to adopt even more complex systems, including models that mimic the way large language models process text, and systems that analyze the relationships between different companies as if they were nodes in a giant, shifting web. The variety of tools has exploded, with researchers now combining price data with news articles, social media sentiment, and even supply chain connections to make their predictions.
However, the authors discovered that this explosion in complexity has not been matched by a similar increase in reliability. When they audited the studies for reproducibility—the ability for other scientists to take the same data and code and get the same results—they found a startling shortage. Only about 7.6 percent of the 685 studies provided the actual computer code needed to run the models. The vast majority of researchers published their findings without sharing the tools used to generate them. This means that most of the reported successes cannot be verified. Furthermore, the researchers identified five specific ways that studies often inadvertently introduce errors, even without intending to. One common error involves using information from the future to train the model. For example, if a researcher normalizes data using the highest and lowest prices of the entire ten-year period before splitting the data into training and testing sets, the model effectively sees the future during its training phase. This creates an illusion of high accuracy that collapses the moment the model is used on new, unseen data.
The review also highlighted a disconnect between how models are tested in the lab and how they perform in the real market. Many studies report impressive accuracy rates, sometimes claiming to predict the direction of a stock correctly more than 90 percent of the time. Yet, these studies often ignore the costs of trading, such as fees and the difference between the price a trader wants to pay and the price they actually get. When the researchers looked for studies that included these real-world frictions, they found very few. Additionally, most papers did not use rigorous statistical tests to prove that their results were better than random chance. Instead, they relied on simple comparisons that could easily be misleading. The authors noted that a model might appear to be a genius simply because it is good at repeating yesterday's price, a trick that works well on paper but earns no money in a live market.
Despite these flaws, the researchers did not conclude that artificial intelligence is useless for finance. Instead, they argued that the field is at a turning point. The tools are powerful, but the way they are being used and evaluated is too loose to be trusted. The team proposed a new set of standards to fix these issues. They called for researchers to share their code and data openly, to test their models using rolling time windows that prevent future data from leaking in, and to include the costs of trading in their simulations. They also insisted that claims of success must be backed by formal statistical proof, not just by showing a high number on a graph. Without these changes, the authors warn, the field will continue to produce flashy results that fail to translate into real economic value. The path forward requires a shift from chasing the most complex algorithm to building the most honest and verifiable one.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.