The MIT Lab Building an AI That Can Predict Stock Crashes With 87% Accuracy

Hurricane Beryl was raging across the Caribbean in early July 2024, with winds as high as 165 miles per hour. Mexico was identified as the most likely landfall location by forecasters at some of the most sophisticated meteorological agencies in Europe, using models on enormous supercomputers that consumed enormous amounts of processing power. GraphCast, a smaller experimental system developed by Google’s DeepMind that could be trained on a laptop, disagreed. Texas was mentioned. GraphCast had been correct and the supercomputers had been incorrect when Beryl hit Matagorda Bay on July 8. If a pattern-recognition AI can read the atmosphere more accurately than the world’s best physics-based models, what might it do with a stock market? This was the obvious question that followed, at least for those who track both weather systems and financial markets for a living.

It is no longer a theoretical question. Deep learning models were tested against 40 years of stock market data by MIT’s Sloan School of Management. The results are worth carefully considering, not because they are consistently comforting but rather because they are genuinely mixed in significant ways. About 80% of minor market corrections—the 5–10% declines that typically frighten retail investors but don’t actually represent crashes—were accurately identified by the models. Accuracy dropped to 37% during significant declines, such as those that surpass 20% and cause entire portfolios to be rearranged. That is preferable to chance. Additionally, you wouldn’t want to stake a pension fund solely on it.
The true story lies in the difference between those two figures. Unusual spikes in trading volume, changes in the VIX, momentum reversals, and variations in the put-call ratio are examples of recognizable technical patterns that are typically followed by minor corrections. AI models that have been trained on decades of historical data are truly adept at reading these kinds of signals. Major collisions are not the same. A pandemic, a bank run that escalated more quickly than any model predicted, or a geopolitical shock that drastically altered global supply chains overnight are examples of events that typically set them off. Some AI-driven hedge funds correctly identified early warning signs of the 2008 financial crisis and profited handsomely from shorting financial stocks while the majority of the market remained optimistic. However, when the same systems attempted to model the recovery following the COVID crash in March 2020, they encountered difficulties. A market that crashed in four weeks and then roared back on a wave of government stimulus that the models had no framework to predict was unlike any recession in history.
An analogy akin to the GraphCast comparison has been used by TradeSmith, a financial analytics company with a staff of 74 researchers and an annual budget of $8 million. With reported annualized returns of 374 percent over the last five years—a figure that accounts for pandemics, geopolitical unrest, and the 2025 tech selloff—their “Super AI” system claims 85 percent backtested accuracy in predicting stock prices up to 21 trading days out. It’s an astounding figure, and it merits the suspicion that astounding figures in finance typically arouse. Live performance and backtested accuracy are two different things. Markets are competitive systems where information is priced in quickly, and an advantage found in past data tends to vanish once enough people are aware of it and take action.
Perhaps the most instructive recent example of how these systems can go wrong in unexpected ways is the 2025 AI selloff. A wave of skepticism led to a correction following a notable surge in AI-related stocks through 2024. AI-generated financial reports pointing out overpriced tech stocks contributed to a mass selloff that momentarily cost Nvidia more than $200 billion in market capitalization in a single week. The AI models were simultaneously executing exits, reading the same signals, and coming to the same conclusions. The event turned out to be the prediction. This structural risk, which researchers refer to as a self-fulfilling prophecy, increases in importance as AI-driven funds make up a greater portion of overall market activity. About 71% of fund manager calls were matched by AI trading systems, according to a Harvard study published in February 2026. This is impressive, but it also indicates that a significant portion of the market is now making correlated decisions based on correlated inputs.
According to JPMorgan Chase, AI-driven funds adjusted portfolios about twice as quickly during volatility and generated 14.6 percent annualized returns compared to 9.3 percent for human-managed funds between similar periods. These are genuine benefits. However, the same analysis pointed out that AI was unable to predict geopolitical-driven crashes; for example, models trained on economic fundamentals were unable to predict how the 2022 Russia-Ukraine war would affect energy prices and European markets. The training data of a system constructed in 2020 does not show any historical pattern for “major European land war followed by Western energy sanctions”.
The similarities to weather forecasting that permeate all of this are difficult to ignore. Because GraphCast learned from 40 years of real atmospheric outcomes rather than attempting to solve fluid dynamics equations from first principles, it was able to predict Beryl better than the supercomputers. Instead of assuming markets behave in accordance with neat theoretical models, the financial analogy would be a system that learns from 40 years of actual market outcomes. In general, MIT models and systems such as TradeSmith’s are trying to achieve that. The underlying issue is that, in contrast to the atmosphere, the stock market has its own observers who alter their behavior when they are aware that they are being observed or that an AI is making predictions about their future actions. The weather does not read its own forecast and then choose to travel in a different direction.
Alongside the market prediction research, it is worthwhile to read MIT’s separate 2025 enterprise AI report, which revealed that 95% of generative AI pilots at businesses are failing. The report’s main problem is a learning gap: AI performs remarkably well on routine, well-defined tasks but falters on high-stakes, non-routine decisions when the circumstances don’t closely resemble anything in its training data. Almost by definition, a market crash is an unusual occurrence. It is still genuinely unclear whether any AI system can consistently close that gap.