All Systems Will Be Gamed
Why adaptive agents always find the cracks, why equilibrium thinking cannot see them coming, and why the programme that assumes boundaries will break is the one that survives
There is a general rule in social and economic life, and W. Brian Arthur stated it with characteristic bluntness: given any system, people will find a way to exploit it.
Or to say it more succinctly: all systems will be gamed.
This is not a law of physics. It is an observational truism that Arthur raised to the level of principle. Given any governmental system, any regulatory framework, any financial structure, any set of rules, people will find unexpected ways to manipulate it to their advantage. “Show me a 50-foot wall,” said Arizona governor Janet Napolitano, speaking of border security in 2005, “and I’ll show you a 51-foot ladder.”
Foreseeing 51-foot ladders may not be particularly challenging. But anticipating how complex economic systems will be exploited by adaptive agents operating within them is another matter entirely. And the consequences of failing to anticipate it can be catastrophic.
Russia’s 1990 transition from planned socialism to capitalism: a small number of well-positioned players seized control of the state’s newly freed assets, creating an oligarchy overnight. California’s 2000 energy deregulation: a small number of suppliers manipulated the freed market to extract billions. Iceland’s banking system in 2008: a few financial players took control of the state’s banks and used depositors’ assets to speculate in overseas property, running the banks into insolvency.
And then there was Wall Street.
Arthur’s paper, published in Complexity and the Economy in 2014, asked a straightforward question. Given that economics is a sophisticated discipline, given that economists study proposed policy systems in advance, how could these various economic disasters have happened?
His answer was structural, and it strikes at the heart of everything we have discussed in this series.
Equilibrium economics, by its base assumptions, is not primed to look for the exploitation of systems. Economic analysis assumes equilibrium: a condition where no agent has any incentive to diverge from its present behaviour. If the system is in equilibrium, then by definition no one has reason to game it. If the system could be gamed, some agents would be initiating new behaviour, and the system could not have been in equilibrium. The conclusion follows with the force of a syllogism: in an equilibrium framework, exploitative behaviour cannot happen.
It cannot happen. Not “it is unlikely” or “it is unusual.” It is structurally invisible. The framework that is supposed to describe reality has made the most dangerous feature of reality impossible to see.
Arthur drew an analogy to engineering. In structural engineering, there is a rigorous sub-discipline of failure mode analysis. Before a bridge is built, engineers study how similar structures have failed in the past. They stress-test the design. They look for the points of maximum vulnerability. They assume that the structure will be subjected to forces its designers did not anticipate, and they build in margins of safety accordingly.
Economics has no equivalent. There is no failure mode analysis for financial systems. There is no tradition of asking: how will adaptive agents exploit the incentives this system creates? How will the rules be gamed? Where are the cracks that a 51-foot ladder will find? The reason is not that economists are careless. The reason is that the foundational framework assumes the question does not arise.
This is the blind spot. And in 2008, the blind spot swallowed the global economy.
Arthur identified four categories of exploitation that recur across systems. They are worth understanding, because each one played a role in the crisis that was already building when he began writing.
The first is the use of asymmetric information. One party understands the product or the system far better than the other, and uses that understanding to extract profit at the other’s expense. In 2007, Goldman Sachs created a package of mortgage-linked bonds and sold it to clients. But Goldman allowed a prominent hedge fund manager, John Paulson, to help select the bonds for the package, bonds that Paulson privately believed would lose value, and then to bet against the very product Goldman was selling. The package was a synthetic collateralised debt obligation tied to subprime residential mortgages. It was complicated enough that the buyers could not assess what they were really buying. Paulson profited enormously. The clients lost over a billion dollars. The information was asymmetric by design.
The second is gaming performance criteria. When behaviour is measured by strict metrics, agents optimise to the metric rather than to the underlying intention. Before the crisis, credit rating agencies made their models available to the Wall Street firms whose products they were rating. The firms learned to massage the inputs: to structure their securitisations in ways that satisfied the model’s criteria while loading the underlying portfolios with lower-quality assets. The ratings said AAA. The reality was junk. The agents had gamed the measurement system.
The third is gaining control of system elements. Within AIG, the insurance giant, a small group of people in the Financial Products unit managed to take effective control of much of the company’s risk-bearing capacity. The unit, led by Joseph Cassano, sold hundreds of billions of dollars in credit default swaps, essentially insurance contracts on mortgage-backed securities, without putting up meaningful collateral. A unit of roughly four hundred people had placed the entire firm, a company with a trillion dollars in assets, at the mercy of the US housing market. They had captured a system element (AIG’s AAA credit rating and its balance sheet) and used it for purposes its designers had never imagined.
The fourth is using system elements in ways not intended by their designers. The securitisation process itself was designed to distribute risk: to take pools of mortgages, slice them into tranches of varying risk, and sell the tranches to investors with matching appetites. In practice, the process was used not to distribute risk but to manufacture the appearance of safety. Mortgage originators had no incentive to assess creditworthiness because they immediately sold the loans. Investment banks had no incentive to worry about default because they immediately securitised. Rating agencies had no incentive to be rigorous because they were paid by the firms whose products they rated. At every link in the chain, the system’s design was used in a way its architects had not intended, by agents whose incentives pointed in directions the designers had not foreseen.
Four categories. Four ways that adaptive agents find cracks in any system built on fixed rules. Arthur’s point was not that these agents were unusually greedy or corrupt. His point was that exploitation is not an aberration. It is the natural, inevitable consequence of placing adaptive agents inside a rule-based system and assuming they will behave as the rules intend.
Now translate Arthur’s framework into the language we have built across this series.
Every actor in the 2008 crisis was making a prediction.
The mortgage originators predicted that housing prices would continue to rise, or more precisely, they predicted that it didn’t matter because someone else would bear the risk. The investment banks predicted that the securitised products they created were safe, or more precisely, they predicted that the ratings agencies’ models were correct. The ratings agencies predicted that historical default rates on mortgages were a reliable guide to the future. The investors who bought the AAA-rated tranches predicted that the rating meant what it had always meant: near-zero probability of loss. And AIG’s Financial Products unit predicted that the credit default swaps they had sold would never be called upon, that the housing market could not decline far enough to trigger the insurance they had written.
Joseph Cassano, the head of AIG Financial Products, said it plainly in August 2007: “It is hard for us, and without being flippant, to even see a scenario within any kind of realm of reason that would see us losing one dollar in any of those transactions.”
Not one dollar. In a portfolio of over $500 billion in notional credit default swap exposure.
This was not stupidity. It was the terminal stage of predictive strategy crowding. Every agent in the system was making a variation of the same prediction: housing is safe, mortgages are sound, the structure will hold. The prediction had become so universal that it had achieved the status of fact. It was, in the language of the El Farol problem from Article 2, an ecology of predictions that had become dangerously homogeneous. Everyone was going to the bar.
And as Arthur proved in 1994, any prediction that becomes universal in a reflexive system is guaranteed to fail. When the prediction fails, it fails catastrophically, because the unanimity of the bet means there is no one on the other side.
The collapse began slowly, as collapses do.
In late 2006, the rate of serious mortgage delinquencies began to tick upward. In early 2007, two Bear Stearns hedge funds that held large positions in subprime mortgage-backed securities reported heavy losses. In August 2007, Goldman Sachs began making collateral calls on AIG: we think the value of the bonds you’ve insured has declined; put up cash to cover the difference.
AIG resisted. Cassano disputed the valuations. The firm posted some collateral but contested the amounts. For months, the dispute remained a private negotiation between counterparties. The ecology of predictions was fraying at the edges, but the core prediction (housing is fundamentally sound, the system is stable) still held.
Then the boundary broke.
In September 2008, the sequence of events compressed into days. Lehman Brothers filed for bankruptcy on September 15. The next day, the rating agencies downgraded AIG’s credit rating. The downgrade triggered contractual clauses in AIG’s credit default swaps that required the firm to post an additional $14.5 billion in collateral immediately. AIG did not have it. The firm that had insured half a trillion dollars in mortgage risk could not meet a margin call.
The US government stepped in with an $85 billion emergency loan on September 16, 2008, taking 79.9% ownership of the company. It would eventually commit $182 billion to prevent AIG’s collapse. The money flowed through AIG to its counterparties: $14 billion to Goldman Sachs alone, at one hundred cents on the dollar, for insurance contracts that AIG could never have paid on its own.
No single actor caused the crisis. No one decided to destroy the global financial system. What happened was emergent. Mortgage originators, investment banks, rating agencies, insurers, and yield-hungry investors each made locally rational predictions about risk. Each gamed the system in one of Arthur’s four categories: asymmetric information, gaming criteria, capturing system elements, using the system in unintended ways. The global outcome, a cascading collapse that wiped out $2 trillion in bank capital worldwide, was a property of the system, not of any individual within it.
It was a complexity event. And equilibrium economics, by construction, could not see it coming.
In the same year that the equilibrium models were saying the world was fine, diversified systematic trend following programmes captured some of the largest moves in a generation.
They did not predict the crisis. No one predicted it. What they did was something structurally different: they assumed that boundaries would break.
This is the critical connection between Arthur’s “all systems will be gamed” and the practice of trend following. If you accept Arthur’s premise (that adaptive agents will always find the cracks in any rule-based system), and if you accept the El Farol insight (that predictive strategies will always crowd and fail in reflexive markets), and if you accept increasing returns (that positive feedback will amplify small moves into large ones), then you arrive at a single, unavoidable conclusion: boundaries will break. The only questions are which boundary, when, and in which direction.
A predictive strategy tries to answer those questions. It builds a model. It identifies the boundary it believes will break, and positions accordingly. This can be spectacularly profitable when the prediction is correct and early. But it is subject to every vulnerability Arthur described: the prediction can be wrong, the timing can be off, the ecology of predictions can crowd the same trade, and the agent can be on the losing side of a complexity event that no model anticipated.
A responsive strategy does not answer those questions. It accepts that the questions are unanswerable. It positions across dozens of uncorrelated markets, each containing its own ecology of predictions, its own gaming dynamics, its own compressed boundaries. When a boundary breaks somewhere in the world, the responsive strategy detects the directional move and follows it. When the move reverses, it cuts. It does not need to know why the boundary broke. It does not need to have predicted the crisis. It needs only to be present when the stored energy releases.
In 2008, the stored energy was immense. Commodities trended violently: oil surged above $140 before collapsing. Bonds rallied as capital fled to safety. The US dollar strengthened as global deleveraging unwound carry trades. Equity indices fell with persistence and momentum. Each of these moves was the structural consequence of prediction failure: the ecology of predictions about housing, about risk, about stability, had collapsed, and the energy that had been compressed behind those predictions released into trends.
The trend following programmes were there. Not because they had seen it coming. Because they were built to be there whenever it came.
Arthur called for economics to develop a failure mode analysis tradition, parallelling the disciplines that protect lives in structural engineering and aviation.
I have spent the past several years thinking about what that would look like in practice, and the result is Carved by Impossibility (forthcoming 2026). The title comes from a principle I have come to believe is the most important in portfolio construction: robust structure is not designed. It is what remains after everything fragile has been eliminated.
Consider what would happen if you took Arthur’s framework seriously as a portfolio manager.
You would not assume that the system is in equilibrium. You would assume it is being gamed, everywhere, all the time, by agents whose behaviour you cannot predict. You would not assume that your models capture reality. You would assume that reality will produce outcomes your models have not anticipated. You would not concentrate your capital in a single prediction about which boundary will break. You would distribute your capital across many boundaries, in many markets, accepting that some will hold and some will shatter, and that you cannot know in advance which is which.
You would not predict. You would respond.
This is not abstract philosophy. This is the architecture of a diversified systematic trend following programme. And it is, I believe, the only architecture that takes Arthur’s insight to its logical conclusion: if all systems will be gamed, and if the gaming will produce cascading failures that no model can foresee, then the only durable strategy is one that does not require foresight. One that is already positioned, across enough markets, with enough breadth, that whenever the next crack opens, the programme is there to capture the energy that escapes.
Arthur used a beautiful image in his paper. He compared the exploitation of systems to the way water finds cracks in a dam. The exploitation emerges from the interaction between adaptive behaviour and fixed structure. The water does not plan. It does not model the dam’s weaknesses. It simply flows, applying pressure continuously, until the structure yields.
Markets are water.
Every fixed structure in finance (a currency peg, a risk model, a regulatory framework, a ratings methodology, an implicit government guarantee) is a dam. And the adaptive agents within the market are water: probing, pressing, flowing around constraints, exploiting incentives, until the structure yields. It always yields. Not because the agents are malicious, but because they are adaptive. Because they are, in Arthur’s language, inductively probing the system to find out what works.
The 2008 crisis was water finding cracks. The Swiss franc de-peg was water finding cracks. Volmageddon was water finding cracks. Every regime break, every lock-in failure, every compressed boundary that finally shatters is the same dynamic: adaptive agents exploiting a fixed structure until it gives way.
A predictive strategy tries to identify which dam will break. Sometimes it is right. When it is wrong, or when the dam holds longer than the prediction can afford to wait, the strategy fails.
A responsive strategy does not try to identify the dam. It monitors the water. When the flow changes, when the pressure shifts, when the energy begins to release, the responsive strategy aligns with the flow and follows it.
Niels Kaastrup-Larsen and I discuss this at length in Trend Following Manifesto (forthcoming 2026): the portfolio is not a collection of opinions about which systems will fail. It is an array of sensors positioned at dozens of different dams, each one monitoring the flow, each one ready to respond when the water finds the crack. The programme does not need to know the dam’s engineering. It needs only to detect the change in flow.
Every few years, a crisis erupts that the equilibrium models did not predict. Every time, there is a scramble to explain how the models missed it: the parameters were wrong, the tail risk was underestimated, the correlations broke down. And every time, the models are patched, the parameters are updated, and the same fundamental assumption is reinstated: the system is in equilibrium, agents behave rationally, exploitation is not a concern.
Arthur’s work tells us something different. The models did not fail because the parameters were wrong. They failed because the framework is structurally incapable of seeing what it needs to see. Equilibrium is a condition where no agent has incentive to change. Reality is a condition where every agent is constantly probing for advantage. These are not the same world.
All systems will be gamed. All predictions will eventually crowd. All compressed boundaries will eventually break. These are not risks that can be modelled away. They are permanent features of a market populated by adaptive agents operating in a reflexive system.
The conventional approach is to try to predict which system will fail, and to position before it does. This is a predictive strategy. It lives inside the ecology. It is subject to crowding, to reflexive self-destruction, to all the dynamics that Arthur and the El Farol problem describe.
The alternative is to accept that systems will fail, that you cannot know which ones, and to build an architecture that is permanently positioned to respond when they do. This is a responsive strategy. It sits outside the ecology. It harvests the energy that prediction failure releases.
In The Fractals of Finance, I wrote that boundaries compress and then release. In Complex Adaptive Markets (arriving soon), I show how the living systems inside markets produce these dynamics. In Carved by Impossibility (which will be released 2026), I argue that the robust portfolio is not the one with the best predictions. It is the one that survives after every fragile prediction has been shattered.
Arthur told us the rule. All systems will be gamed. The water will find the cracks. The boundary will break.
The only question is whether your programme is built to respond when it does.
Follow the break.
Next in the series: “The Physicist Who Beat the House Twice.” J. Doyne Farmer: from roulette wheels to Wall Street to simulating entire economies, and what his journey tells us about the science of responsive strategy.
Want the theoretical foundation for why trend following works?
The Fractals of Finance: Determinism, Adaptation and the Geometry of Markets
The book explores the full architecture of feedback, fat tails, and fractal structure in financial markets, and what it means for how we trade, invest, and understand risk.
Available now on Amazon in paperback, hardcover, and Kindle.
Want a practical field manual for trading trends and capturing outliers?
The Aussie Turtles Trend Following Guide: A Field Manual for Hunting Outliers adapts the timeless principles of the original Turtle traders into a systematic, rules-based approach for modern markets. Co-authored with Adam Havryliv.
Available now on Amazon in paperback, hardcover, and Kindle.