The Model Has Never Seen a War Start on a Tuesday Morning
By R Tamara de Silva
The promise that a quantitative model can replace human judgment in financial markets has now failed twice in the span of a single career. The first failure was Value at Risk. The second, still unfolding, is artificial intelligence. The failures are not coincidental. They are structurally identical. Both VaR and AI are averaging engines trained on historical data, and both go blind at exactly the moment that matters most: when something happens that has never happened before.
This is not an argument against artificial intelligence. AI is an extraordinary tool, and I use it daily. This is an argument against a specific and recurring fantasy: that a model, any model, can replace the human being sitting in front of the screen. It cannot. What it can do, and do well, is make that human being significantly better. The distinction is the whole point of this essay.
I have some experience with this problem. I spent years on the floor. I have represented traders and funds on both sides of the regulatory apparatus. And in August 2009, a few months after the Credit Crisis had vaporized the balance sheets of firms that had staked their survival on VaR, I wrote a piece on this site called “Don’t Tell Me You Never Heard of Goldman Sachs?!” that tried to explain what had just happened and why. Goldman’s VaR had jumped 135 percent between 2001 and 2006. Lehman, Morgan Stanley, and every other firm with a serious trading desk had hired armies of mathematicians from New Delhi to MIT to build increasingly sophisticated VaR variants. They believed that hyper-diversification across global markets made the chance of everything going wrong at once statistically negligible. Then everything went wrong at once.
What concerns me now is that the same logic, wearing new clothes, is being sold again.
Taleb Was Right
Nassim Taleb saw this coming before anyone else, and he said so publicly, repeatedly, and at personal cost to his reputation among the quantitative establishment.
In April 1997, in the pages of Derivatives Strategy magazine, Taleb debated Philippe Jorion, the most credible academic defender of VaR, and called the entire framework charlatanism. Not flawed. Not in need of refinement. Charlatanism. His argument was precise. VaR had already been falsified by real markets in 1985, 1987, 1989, 1991, 1992, 1994, and 1995, a record of failure that would have killed any hypothesis in a scientific discipline that took its own standards seriously. The model ignored 2,500 years of accumulated trader experience in favor of untested mathematics built by people who had never traded. And at its core, VaR rested on a claim that no honest statistician should make: that the probability of rare, extreme events can be reliably estimated from historical data. It cannot.
Taleb drew a distinction that matters here. Traders, he said, are clinical researchers. They work on real patients in real time. Modelers run simulations on imagined ones. The difference is not merely philosophical. It is the difference between knowing that prices can do something on a given morning that no dataset has ever recorded and believing that the dataset is the world. That distinction is the whole argument of this essay, and it applies to AI with the same force it applied to VaR.
Twelve years later, in 2009, Taleb testified before Congress and asked for VaR to be banned outright. By then the evidence was overwhelming. He has not changed his position since, and history has not given him a reason to.
What Fat Tails Actually Mean
The core of the problem is mathematical, but the intuition is simple enough.
Most quantitative risk models assume that market returns follow a normal distribution, the familiar bell curve taught in every introductory statistics class. Under a normal distribution, very large moves are vanishingly rare. A twenty-standard-deviation event, for example, is so improbable that the universe is not old enough for one to have occurred.
The October 1987 crash, in which the S&P fell more than 20 percent in a single session, was a twenty-standard-deviation event. It happened anyway. So did the Asian crisis in 1997, the Russian default and LTCM collapse in 1998, the dot-com implosion in 2000, and the Credit Crisis of 2008. So did everything that followed.
Real markets do not have normal distributions. They have fat tails. The term refers to the shape of the probability curve: the “tails” at either extreme, where the largest gains and losses live, are much fatter than a bell curve predicts. Extreme events are not rare aberrations. They are a defining and recurring feature of how markets actually behave.
This is not an academic distinction. When you build a model on the assumption that returns are normally distributed, you have not modeled the real market. You have modeled a fictional one that shares the same average behavior on quiet days. The model will look correct for years, precisely because nothing has happened. Then something will happen, and the model will not merely be wrong. It will have nothing to say at all. The event is outside the distribution it knows.
Hank Paulson, of all people, conceded this in November 2006. Asked about the spike in Goldman’s VaR, he acknowledged that VaR “always assumes that the future is going to be like the past,” and that in a shock, “things you hope wouldn’t correlate are going to correlate.” That was the chief executive of the most sophisticated trading firm in the world admitting that his risk model was structurally blind to the events that would actually kill it. David Einhorn put it more plainly in his 2008 debate with Aaron Brown in GARP Risk Review: VaR is an airbag that works all the time, except when you have a car accident.
The Crises Are Getting More Frequent, Not Less
This is not an occasional problem. It is now the dominant pattern.
The IMF’s systemic banking crisis database, compiled by Luc Laeven and Fabián Valencia, documents 151 systemic banking crises between 1970 and 2019, alongside 414 currency crises and 200 sovereign debt crises. Work by Michael Bordo and his coauthors found that the frequency of financial crises since 1973 has been roughly double the rate observed under Bretton Woods and the classical gold standard, rivaled only by the 1920s and 1930s. These are not my numbers. They are the IMF’s.
Since I wrote on VaR in 2009, the list has only grown. The May 2010 Flash Crash. The European sovereign debt crisis from 2010 to 2012. The Swiss franc unpegging in January 2015. The Chinese equity collapse of mid-2015. Brexit in 2016. The volatility implosion of February 2018, when the VIX short trade blew up in a single afternoon. The COVID liquidity crisis of March 2020, when the Treasury market itself briefly stopped functioning. The UK gilt crisis and LDI cascade of September 2022. The regional banking failures and Credit Suisse collapse in March 2023. The Trump tariff announcements in April 2025 that wiped roughly six trillion dollars of equity value in two sessions. And the US-Iran war that began on February 28, 2026, which sent Brent crude from roughly $72 to $112 in a matter of weeks. The IEA called it the largest supply disruption in the history of the global oil market.
That is not a list of anomalies. It is the environment. If “one or two crises per decade” was ever an accurate description of financial markets, it has not been one for a long time. We are at roughly one significant dislocation per year, sometimes more. Each one is a fat-tail event. Each one is supposed to occur, under the assumptions embedded in conventional risk models, once every several thousand years. They occur instead with a regularity that has become routine to anyone who actually trades for a living.
AI Has the Same Architecture and the Same Blind Spot
A large language model is an averaging engine. It learns from text and data that already exists. It reflects, with remarkable fluency, the consensus pattern of what has been written and traded before. This makes it useful on an ordinary day. But on an ordinary day, prediction markets and the consensus tape have already priced in everything publicly available before you opened your laptop. Same feeds. Same data. Same conclusions. Trading the consensus pattern is not edge. It is confirmation with better processing speed.
The real question is what happens on the days that are not ordinary.
No model saw the Iran war coming. No model saw the tariff wipeout coming. More to the point, no model knew what to do after either one happened, because no model had ever encountered them. This is not a flaw that can be fixed with more data or more parameters. A model trained on the past cannot anticipate a discontinuity from the past. That is not a limitation of the current technology. It is a property of what averaging means. The same mathematical architecture that makes the model useful on Monday makes it worthless on the Tuesday morning a war breaks out.
Taleb said this about VaR in 1997. Paulson conceded it about VaR in 2006. The Credit Crisis proved it in 2008. Every fat-tail event since has proved it again. The argument does not change because it cannot change. An averaging engine, by construction, cannot price what it has never seen.
Where AI Actually Earns Its Keep
None of this means AI is useless in markets. It means AI is useless in the way it is being marketed.
The actual capabilities are formidable. AI can process earnings transcripts, regulatory filings, and macro data releases faster and more comprehensively than any human analyst. It can scan for correlations across asset classes that would take a research team days to identify. It can model scenario trees, run Monte Carlo simulations, and back-test a strategy against decades of data in minutes. It can monitor position risk in real time and flag concentration problems before they become losses. It can draft compliance documentation, summarize legal developments, and keep a trader current across multiple jurisdictions simultaneously. These are not trivial contributions. For a solo practitioner or a small fund, they represent a genuine force multiplier.
I trade my own book and I use AI as a risk manager. Not as an oracle. The distinction matters. AI is a superb tool for stress-testing a human thesis. You bring the read. The model pressure-tests it. It finds the contradictions you missed. It surfaces historical analogues you had not considered. It tells you what consensus has already priced in, so you know whether your trade is actually a trade or just an expensive agreement with the tape. It runs the math on your scenarios so you can size positions correctly.
The point is that AI makes a good trader better. It does not make a non-trader into a trader. The thesis itself, the read on what the world is becoming rather than what it has been, is still yours. No human logic going in, no edge coming out. The model processes the data you give it. The part where the money actually lives, the judgment about direction and timing that the data does not yet contain, is on you. It has always been on you. AI will not replace that judgment. It will, if used honestly, sharpen it. That is not a small thing. But it is a different thing from what is being promised.
The Same Seduction in a New Box
The pattern repeats because the seduction repeats. In 2006, the seduction was that mathematics had finally caught up to risk. Seventeen years later, in 2026, the seduction is that the machine has finally caught up to judgment. Neither was true. Both arrived at exactly the moment the people making the claim most needed to believe it.
AI will not replace traders. It will replace traders who refuse to use it. But the tool is not the trade. The trade is the read, and the read belongs to the person who has been in the room when the world changes.
The model has never seen a war start on a Tuesday morning.
You have.
R Tamara de Silva
Sources
Nassim Nicholas Taleb, “Against Value at Risk,” in “The Jorion-Taleb Debate,” Derivatives Strategy, April 1997.
Nassim Nicholas Taleb, The Black Swan: The Impact of the Highly Improbable (Random House, 2007).
Nassim Nicholas Taleb, Testimony before the Subcommittee on Investigations and Oversight, Committee on Science and Technology, U.S. House of Representatives, September 10, 2009. The Risks of Financial Modeling: VaR and the Economic Meltdown, Serial No. 111–25.
Luc Laeven and Fabián Valencia, “Systemic Banking Crises Database II,” IMF Economic Review, Vol. 61, No. 2 (2013), pp. 225–270.
Wee Chian Koh, M. Ayhan Kose, Peter S. Nagle, Franziska L. Ohnsorge, and Naotaka Sugawara, “A New Comprehensive Database of Financial Crises: Identification, Frequency, and Duration,” Economic Modelling, Vol. 109 (2022).