SDSU researchers build AI to predict stock market trends and volatility

Traders have spent decades split between two camps: chart watchers analyzing moving averages, and researchers digging through financial filings. The problem is that neither approach handles sudden macro shocks or social media frenzies very well.

Now, researchers from South Dakota State University, working alongside Virginia Tech and Texas A&M University-Corpus Christi, think they have a better framework. Led by SDSU assistant professor Kaiqun Fu and graduate assistant Yangxiao Bai, the team introduced ALERTA-Net, an AI model built to predict both stock price movement and market volatility trends. The preprint is published on arXiv.

Mixing 7 million tweets with Fed data

Traditional technical analysis relies heavily on historical prices, which Fu notes can easily overlook sudden events and create a self-reinforcing echo chamber disconnected from real economic conditions. Fundamental analysis is broader, but often ignores the deep feedback loop between financial markets and the macroeconomy.

Another big limitation in existing predictive models is that they usually try to guess whether a stock goes up or down, while ignoring how big the swing might be. In real trading, magnitude is everything.

To tackle this, the researchers selected 41 blue-chip stocks and pulled three years of price history from Yahoo Finance. Then they brought in macroeconomic records from the Federal Reserve along with search engine traffic.

They also tapped into social sentiment. Fu and Bai gathered more than 7 million tweets from X that mentioned those 41 stocks, analyzing both daily post volume and individual post sentiment. In benchmark tests, ALERTA-Net outperformed established architectures like DP-LSTM as well as standard baselines.

Fu says adding macroeconomic figures significantly boosted the network's predictive edge, and the team is already looking at expanding the model to parse audio and video data next.

Why it matters

If you are an individual investor expecting an off-the-shelf money maker, this is not it. The research is an early academic benchmark tested on 41 highly liquid blue-chip tickers, far removed from an automated live trading setup.

What it signals for the industry, though, is how quant modeling is evolving. Markets are increasingly driven by online sentiment colliding with macro policy shifts. Models that rely purely on candlestick math are getting outmatched by architectures that treat Fed moves, search interest, and social posts as connected signals.

We think the smart move here was training the network on both price direction and volatility magnitude simultaneously. For risk managers and derivatives desks, the size of a swing is often far more critical than just knowing whether the day closes green or red.

The real test will be data contamination. Social media platforms like X are crawling with promotional bots and coordinate campaigns. If this model were deployed live, filtering out weaponized noise will be just as hard as parsing genuine sentiment.