Every morning, financial news pours into investors’ inboxes, phones, and terminals. Some of it feels alarming, some reassuring, and some just noise. The question of whether that daily tone actually moves markets, and whether artificial intelligence can measure that tone reliably, has become one of the more active corners of financial research.
A new paper published in Applied Economics Letters examines exactly this. The authors used a large language model to score more than 29,000 Chinese financial news headlines, translated those scores into variables drawn from prospect theory, and then tested how well those variables tracked daily returns on the Shanghai Composite Index.
The question behind the study
Sentiment analysis in finance is not new. For years, researchers have tried to measure the mood of news coverage using word lists, dictionaries, or older natural language processing tools. These approaches often struggle with context. The word “cut,” for example, might mean something very different in “rate cut” than in “job cuts.”
Large language models, the same family of AI that powers chatbots, are better at reading context. But as Te-Mei Chiang of Asia University in Taiwan and colleagues at Xiamen University Tan Kah Kee College point out, most previous work using these tools has treated sentiment as a simple linear input. The researchers wanted to connect AI-generated sentiment scores to prospect theory, the behavioral framework developed by Daniel Kahneman and Amos Tversky that describes how people actually weigh gains, losses, and probabilities.
Prospect theory has two key ideas relevant here. First, people feel losses more intensely than equivalent gains, and they behave differently in each domain. This is captured in what’s called a value function. Second, people distort probabilities: they tend to overweight rare events and underweight common ones. That distortion is captured in a weighting function.
How the study was built
The research window ran from January 1, 2024 to January 21, 2026. The authors pulled 29,077 headlines from “Lujiazui Financial Breakfast,” a daily Chinese financial news digest sourced through the Wind database. They also collected daily returns for the Shanghai Composite Index and control variables such as market volatility.
To score each headline, they used DeepSeek, a Chinese large language model. The prompt asked the model to act as an economist and assign a score between -5.0 and +5.0 to each headline, returning the result as structured data. To check whether DeepSeek’s scores were reliable, the team compared them to scores from a second model, Kimi, over two separate time periods. The correlation between the two models’ scores was 0.88 and 0.87, suggesting the sentiment readings were reasonably consistent across systems.
The next step was translating those raw scores into prospect theory variables. Positive scores were fed into the “gain” side of the value function, where investors are treated as risk-averse. Negative scores went into the “loss” side, where investors are treated as risk-seeking. The authors used parameter values from Tversky and Kahneman’s 1992 paper on cumulative prospect theory, standard reference numbers in the field.
For the weighting function, they binned sentiment scores into intervals of 0.5, calculated how often each bin actually appeared in the data, and used those frequencies to build subjective probability weights. Combining the value function and the weighting function produced a “total subjective value” for each day.
What the analysis found
The core variables turned out to be highly correlated with one another, with correlations reaching 0.99 between some sentiment measures. That kind of overlap creates a statistical problem called multicollinearity, which can make regression coefficients unreliable. To handle this, the researchers used Lasso regression, a method that automatically shrinks or drops redundant variables.
After the Lasso procedure, two variables emerged as the meaningful predictors of Shanghai Composite returns: the total subjective value (the combined value-and-weight measure) and the weighting function on its own.
The total subjective value had a positive and statistically significant relationship with daily returns. In plain terms, when the AI-derived sentiment, filtered through prospect theory, pointed toward gains, the index tended to rise. The weighting function, taken alone, had a significant negative relationship with returns. The authors interpret this pattern as consistent with what prospect theory predicts: investors’ tendency to overweight unlikely outcomes and underweight common ones shows up in how prices move.
To check that these results weren’t an artifact of the specific AI model used, the team re-ran the analysis substituting Kimi’s sentiment scores for DeepSeek’s. The core findings held.
Important caveats
The authors are direct about the limitations of their model. Several standard statistical assumptions were violated. A Breusch-Pagan/White test indicated heteroscedasticity, meaning the variability of the errors was not constant. Even after Lasso, the average variance inflation factor for the sentiment variables was 41.68, well above the threshold typically considered acceptable. The residuals were not normally distributed, and no autocorrelation test was performed.
The R-squared from the Lasso model was 0.0096, meaning the sentiment variables collectively explained less than one percent of daily return variation. That is not unusual in daily stock return prediction, where noise dominates, but it is a reminder that “significant” here refers to statistical detection of a pattern, not to a model that would let anyone reliably forecast daily index moves.
The study is also limited to a single news source, a single index, and a roughly two-year window that includes an unusual period in Chinese markets. The authors flag broader news sources, additional market cycles, and refined prompts as areas for further work.
Why the framing matters
What the researchers argue is distinctive about their approach is the pairing of AI-generated sentiment with a behavioral-finance structure rather than a simple positive-or-negative label. Instead of asking “was today’s news good or bad?”, the model asks how investors are likely to weigh that news given known psychological patterns around gains, losses, and probability distortion.
The authors suggest the resulting indicators could inform both individual investor decisions and market oversight by regulators, and that the methodology, including the specific zero-shot prompt used to elicit scores from DeepSeek, is reproducible enough for other researchers to build on. Whether the approach generalizes beyond the Shanghai Composite and this particular news feed is one of the open questions they flag for follow-up work.
