A couple of fascinating new working papers from the Bank for International Settlements (BIS) suggest we’re on the cusp of exactly that. Researchers are pitting new-school machine learning techniques against old-school models in the ultimate cage match for financial stress prediction. And frankly, the results are looking a bit one-sided. It seems the secret sauce isn’t just about crunching numbers better; it’s about teaching machines to read the room.
The Old Guard and the New Contender
For years, the go-to method for economic forecasting has been autoregressive models. Think of these as diligent, if slightly unimaginative, historians. They look at past data for a specific indicator—say, a bond yield spread—and use its own history to predict its future. It’s a bit like driving by only looking in the rear-view mirror. It works fine on a straight road, but it’s utterly useless for spotting a sharp turn ahead.
These models are clean, theoretically sound, and easy to interpret. But their great weakness is their reliance on a narrow set of pre-defined relationships. They struggle with the sheer messiness of modern finance, where fear, sentiment, and a tweet from the right (or wrong) person can move markets faster than any quarterly report. This is where the AI-powered new school wades in.
Instead of just looking at one data point’s history, machine learning models are designed to find complex, non-linear patterns across a vast sea of information. They are the ultimate data detectives, sifting through mountains of evidence to find clues that human analysts might miss.
Case Study One: The Random Forest That Sees the Wood for the Trees
In one corner, we have the work of Aldasoro et al. (2025), detailed in a recent column on CEPR. They decided to build a risk assessment algorithm using a technique called a “random forest.” The name might sound like something out of a fantasy novel, but the concept is brilliantly simple. Instead of relying on one single, monolithic model (one giant, wise old oak tree), a random forest builds hundreds or even thousands of smaller, simpler “decision trees.” Each tree gets a slightly different view of the data and makes its own prediction. The final forecast is then decided by a majority vote, or an average, across all the trees in the forest.
It’s like asking a huge committee of experts for their opinion instead of just one. This approach makes the model incredibly robust and less likely to be thrown off by a single weird piece of data. Aldasoro and his team used this framework to forecast market condition indices across Treasury, foreign exchange, and money markets.
So, how did it fare against the traditional autoregressive models? It wasn’t even close. The paper reports that the random forest models reduced quantile losses by 15-20% at forecasting horizons of up to 12 months. In plain English, that means they were significantly better at predicting extreme events—the so-called “tail risks” that are the stuff of financial nightmares. They weren’t just more accurate on average; they were much better at spotting the really dangerous stuff on the horizon. This isn’t just an incremental improvement; it’s a categorical leap in our ability to perform meaningful AI market analysis.
Case Study Two: Teaching an AI to Read the News
But what if the most important clues aren’t in the numbers at all? That’s the question tackled by the second BIS paper from Aquilina et al. (2025). Their approach is, to my mind, even more interesting because it acknowledges a fundamental truth: markets are driven by people, and people talk. A lot.
Their model combines two powerful AI techniques:
* Recurrent Neural Networks (RNNs): These are great at understanding sequences and time-series data, making them a natural fit for analysing financial metrics that evolve over time.
* Large Language Models (LLMs): This is the same family of technology that powers things like ChatGPT. Aquilina’s team used an LLM to read and interpret the text from thousands of financial news articles.
This hybrid approach creates a model that can listen to the market’s quantitative and qualitative signals. It’s the weather forecasting equivalent of having both a barometer and a system that analyses every weather report, news broadcast, and social media post about an approaching storm. One tells you the conditions are changing; the other tells you why and how people are reacting.
The team set their model a specific task: to predict deviations in the covered interest parity (CIP) for the euro-yen exchange rate. This is a bit of a niche indicator, but it’s a brilliant barometer for systemic stress, particularly around US dollar funding, which is the lifeblood of the global financial system. When this indicator goes haywire, it’s a sign that something is seriously wrong in the plumbing.
The results were, frankly, stunning. The model successfully predicted episodes of market dysfunction up to 60 working days in advance. Most impressively, using only training data from before 2021, the model managed to identify the build-up of risks that preceded the March 2023 banking turmoil, which saw the collapse of Silicon Valley Bank and the emergency rescue of Credit Suisse. It was, in effect, flagging a high probability of a dollar funding squeeze weeks before it actually happened, all by cross-referencing market data with the context it was gleaning from financial news.
Making the ‘Black Box’ Transparent
One of the longest-standing criticisms of using AI in high-stakes fields like finance is the “black box” problem. Sure, the AI might give you the right answer, but if you don’t know why it reached that conclusion, can you really trust it to make multi-billion-pound decisions? It’s a fair question. A forecast is useless to a policymaker if it comes with no explanation.
This is where another clever technique called Shapley values comes in. Borrowed from game theory, this method allows analysts to peek inside the model and see exactly how much each individual input variable contributed to the final prediction.
Following the March 2023 banking turmoil, the researchers used Shapley values to Eexamine their model’s “thinking.” They found that whilst the model was initially focused on things like credit default swaps, its attention dynamically shifted. As the crisis unfolded, the textual analysis of news headlines about “deposit outflows” and “bank runs” became a far more important driver of its high-stress prediction. This provides an audit trail, showing policymakers not just what the model is seeing, but how its assessment is evolving in real-time. It transforms the AI from an inscrutable oracle into a transparent and accountable co-pilot.
What Does This Mean for You?
So, why does any of this matter outside the halls of central banks?
* For Investors: These tools won’t give you stock tips, but they represent a massive leap forward in risk management. Improved financial stress prediction means that the indices and warnings that fund managers and institutional investors rely on will become far more reliable. It could lead to better-timed defensive portfolio adjustments, reducing the impact of market shocks. Over time, as these tools become more widespread, we might see a slight reduction in the kind of ‘panic selling’ that exacerbates crises, as the market becomes better at pricing in risk ahead of time.
* For Policymakers: This is a potential game-changer. The ability to spot a brewing crisis 60 days out, rather than reacting to it in real-time, is the difference between preventative medicine and emergency surgery. It gives regulators precious time to take pre-emptive action, like bolstering liquidity facilities or stress-testing specific institutions, before the fire starts to spread. Accurate economic forecasting, powered by these new risk assessment algorithms, could be our best defence against the next global financial crisis.
The work from the BIS shows we’re moving from a world of simple gauges to one of holistic surveillance. By combining the cold, hard numbers with the nuanced, emotional narrative of the market, AI market analysis is finally starting to provide a picture of reality that is as complex and interconnected as the financial system itself.
Of course, no model will ever be a perfect crystal ball. The future will always hold surprises. But for the first time, it feels like we’re building a system that can at least spot the iceberg before we’re swapping paint with it. The real question now is, when these AI sentinels raise the alarm, will we have the wisdom to listen? What do you think?
