{"slug": "the-next-risk-in-ai-powered-markets-is-not-a-bad-model", "title": "The next risk in AI-powered markets is not a bad model", "summary": "The European Central Bank (ECB) reported in a Research Bulletin published on May 21, 2026, that machine-learning algorithmic trading accounts for 60-70% of equity transaction volumes in the US and other major markets, and simulations showed reinforcement-learning agents coordinated almost completely, withdrawing from a fund even when fundamentals did not justify it. The author, a trader with 15 years of futures experience, argues that the next risk in AI-powered markets is not a single bad model but many good models reaching the same conclusion simultaneously, creating hidden dependence and potential instability.", "body_md": "Artificial intelligence (AI) is becoming part of the machinery of modern markets. Banks, hedge funds and trading firms are using increasingly sophisticated models to identify opportunities, manage risk and execute trades.\n\nMost discussion focuses on one question: What happens when an AI model gets something wrong?\n\nI think there is another risk. What happens when many good models reach the same conclusion at the same time?\n\nAfter roughly 15 years of trading futures, I have learned that markets often become dangerous not because one participant is wrong, but because too many become convinced of the same thing.\n\nAs traders sometimes say: When everyone moves to the same side of the boat, start watching the waterline.\n\nFinancial institutions may believe they are diversified because they use different models, vendors and strategies. But those systems may still be learning from similar information: Price history, volatility, order flow, economic data and news sentiment.\n\nThey may also be optimising toward similar goals. Different models do not need to be identical to produce the same trade.\n\nSeveral independent systems can simultaneously decide to reduce the same exposure, buy the same asset or hedge the same risk. That creates hidden dependence.\n\nA market can appear diversified while many supposedly independent systems are reacting to the same information in similar ways. The real vulnerability may therefore not be one bad model, but the connection between many good ones.\n\nSupervisors have begun to say something similar. In a *Research Bulletin* published on May 21, 2026, four researchers writing for the European Central Bank (ECB) noted that machine-learning algorithmic trading already accounts for between 60 and 70 per cent of equity transaction volumes in the US and other major markets.\n\nTheir simulations found that reinforcement-learning agents coordinated almost completely, withdrawing from a fund even when the fundamentals did not justify it, while large language models were harder to predict. The authors concluded that the architecture of an algorithm is itself a source of financial instability.\n\nCrowded trades are nothing new. Portfolio managers react to the same data.\n\nTraders watch the same levels. Risk desks cut exposure when volatility rises.\n\nTechnology increases the speed of those reactions. AI may also increase their consistency.\n\nHumans disagree, hesitate or sometimes simply choose to do nothing. Machines are less likely to disagree unless that behaviour is deliberately built into them.\n\nIf many models detect the same deterioration in risk and begin reducing exposure simultaneously, individually rational decisions can produce an irrational market outcome.\n\nEach model may be doing exactly what it was designed to do. The market can still become unstable.\n\nA strategy can perform well for years because market conditions remain broadly familiar. That history creates confidence.\n\nBut markets do not owe us tomorrow because something worked yesterday.\n\nOne of the most dangerous moments is not when a model is obviously failing, but when it has accumulated enough successful history that people stop questioning its assumptions.\n\nA calm sea can make almost any captain look skilled.\n\nFor AI-driven markets, the important question should therefore not only be: How accurate is this model? It should also be: What assumptions does this model share with everyone else’s?\n\nThe usual response to model risk is to add more: More data, more parameters, another model, another control.\n\nBut resilience does not always come from adding complexity. Sometimes it comes from removing exposure.\n\nA system may become safer not because it predicts markets more accurately, but because it knows when conditions are unsuitable for taking risk.\n\nThat is one of the hardest lessons to translate from discretionary trading into automation. Finding a trading signal is relatively easy.\n\nTeaching a machine when a valid signal should be ignored is much harder.\n\nExperienced traders know that a setup can look perfectly good and still not deserve capital. The market may be too crowded, liquidity may be changing or correlations may be breaking down.\n\nSometimes the best trade is no trade.\n\nUsing several different AI models does not necessarily provide diversification if all of them react to stress in the same way. Real diversification may require different behaviour, not just different architecture.\n\nWhen should a model refuse to trade? When should it reduce confidence?\n\nWhen should it respond more slowly? And how should it behave when the actions of other models are themselves changing the market?\n\nThat matters because financial markets are not ordinary prediction problems. Predicting the weather does not change the weather.\n\nPredicting a market can change the market because participants act on the prediction.\n\nAs more intelligent systems enter that feedback loop, their collective behaviour may become as important as the quality of any individual model.\n\nFinancial institutions will continue building faster and more powerful models. But I suspect the next competitive advantage will not come only from making machines better at finding trades.\n\nIt will also come from making them better at recognising when everyone else has found the same trade.\n\nMarkets have always punished excessive certainty eventually. AI does not remove that principle.\n\nIt may simply allow certainty to spread faster.\n\nThe smartest trading systems of the next decade may therefore be distinguished not only by how quickly they act, but by something harder to programme: Knowing when not to join the crowd.", "url": "https://wpnews.pro/news/the-next-risk-in-ai-powered-markets-is-not-a-bad-model", "canonical_source": "https://brusselssignal.eu/2026/08/the-next-risk-in-ai-powered-markets-is-not-a-bad-model/", "published_at": "2026-08-25 00:30:11+00:00", "updated_at": "2026-08-25 07:43:49.404912+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-policy", "ai-ethics"], "entities": ["European Central Bank", "ECB"], "alternates": {"html": "https://wpnews.pro/news/the-next-risk-in-ai-powered-markets-is-not-a-bad-model", "markdown": "https://wpnews.pro/news/the-next-risk-in-ai-powered-markets-is-not-a-bad-model.md", "text": "https://wpnews.pro/news/the-next-risk-in-ai-powered-markets-is-not-a-bad-model.txt", "jsonld": "https://wpnews.pro/news/the-next-risk-in-ai-powered-markets-is-not-a-bad-model.jsonld"}}