SMU Assistant Professor Ma Yunshan’s latest research sets out to finetune the way AI predicts stock prices.
By Vince Chong
SMU Office of Research – Large language and deep learning computing models – two systems generally thought of as simply AI – are commonly used these days to quickly process and present massive swathes of information into readable analysis. In financial circles at least, one of the ultimate objectives is to build an algorithm to accurately predict stock movements.
SMU Assistant Professor of Computer Science Ma Yunshan tackles exactly that in one of his latest research projects, which impressively features a leading AI prediction tool to more accurately forecast share prices.
Such work remains at its nascent stage, cautioned the academic who joined SMU in 2025. “At this stage, our work aims to spur on extra research and layer by layer, we might hopefully get to the end one day, but it is a very challenging, very long process,” he told the Office of Research.
“But for this project at least, our results are encouraging.”
The paper in question, Learning to generate explainable stock predictions using self-reflective large language models, represents one more rung on this particular research ladder. Professor Ma cowrote the paper as a postdoctoral research fellow at the National University of Singapore (NUS), along with Kwan Im Thong Hood Cho Temple (KITHCT) Chair Professor Chua Tat Seng and Kelvin Koa, both also of NUS, and Ritchie Ng, Eastspring Investments, Singapore.
As the paper notes, the task of stock prediction remains challenging for LLMs, or large language models, as it “requires the ability to weigh the varying impacts of chaotic social texts on stock prices.”
“The problem gets progressively harder with the introduction of the explanation component, which requires LLMs to explain verbally why certain factors are more important than the others,” it sets out.
“On the other hand, to fine-tune LLMs for such a task, one would need expert-annotated samples of explanation for every stock movement in the training set, which is expensive and impractical to scale.”
To tackle this, Professor Ma and his team harnessed a proprietary summarize-explain-predict (SEP) framework that among other things used a “verbal self-reflective agent” that allows a LLM to “teach” itself how to generate explainable stock predictions.
What this does, simplistically speaking, is to prompt feedback for every instance of prediction that the LLM got wrong, and produce a plan to mitigate this failure when it next makes the same forecast. The LLM then adds such failures to its experiences.
While the project built on and finetuned existing research for the “summarize” and “explain” components of its SEP framework, it is the pioneering study for its way of generating explanations and implementing the “predict” tool for the specific use of stock forecasting.
Encouraging results
In an experiment where the team used their SEP framework to finetune those frameworks currently used by six other AI models to forecast stock prices, the finetuned version outperformed the incumbent iterations. This was based on a set of technical metrics that can be sorted into two groups, one used to evaluate forecasting performance and the other to assess the quality of the generated explanations.
The metrics showed that the framework increased the forecasting performance of the programme to “decisively weigh between news information to make a stock movement prediction.”
They also showed the framework generating better explanations through metrics such as “Global and Industry Factors”, “Contextual Understanding” and “Clarity & Coherence” – a more difficult task for LLMs “as it requires them to not only understand the meaning of natural language texts, but also to correctly reason out their overall impact on the stock price movement.”
So, given the success of the study, did Professor Ma and his team try their prediction framework on real time stock prices? The short answer: No.
“Right now such studies are still at the academic, exploratory stages. While we obtained some kind of performance improvement through our research, extensive engineering and system-level works are required to integrate this core idea into a sophisticated online system,” Professor Ma said with a laugh.
This would require, for one, more computer engineering tests that require tracking not just massive sets of financial and social media data, but also data that changes in real time, like stock market trades.
“[Academics] would need to, for example, collaborate with large financial institutions and social media companies to really push such research but this will not be easy [to set up] as it involves issues such as data privacy and licensing rights,” he added.
“In the real world it’s interactive while research is mostly passive [based on static data] … we and other academics will need to build on these findings to progress.”
Market interest
The paper appears to have been well received since its publication, Professor Ma said, not least because he has been approached to review a growing number of research work on financial forecasting.
“It seems like since our research went out, we have received a lot of emails on the subject, some to seek collaboration, and we also see more and more follow-up papers getting inspiration from our work,” he said.
Further, the research has also attracted the attention of at least two financial institutions – one local, one international – that hope to harness its findings for their business. The sheer prospect of tying up with such institutions, Professor Ma said, is exciting in terms of potential access, even if partially, to a substantial set of dynamic market data.
His team has also collaborated with the Asian Institute of Digital Finance, which is founded by The Monetary Authority of Singapore, Singapore’s National Research Foundation and NUS. This work delves into macro concerns such as improving financial crisis forecasting.
“We utilise what we have done and apply it to macroeconomic indicators such as debt and inflation, or even political conflicts and relations,” Professor Ma said.
“This will better help economies plan ahead and if we contribute to that, however minor, it will be worth it.”
Back to Research@SMU May 2025 Issue
