Haute Lumière · The reading

The market is not a machine, which is why the models keep missing the crash

Finance writes its rules in straight lines and governs a system that runs on loops, and the gap between the two is where the crises live.

The forecast on the table is a week old and she has stopped defending it, which is the harder half of the work.

THE GUARANTEED RETURN

It starts with a spreadsheet, which is how these stories usually start. A young analyst, quick with numbers and willing to sit with a problem for a fortnight, is asked to forecast a quarter for a new fund. He feeds his model everything he can reach: interest rates, inflation, and, because the data was sitting there anyway, projected weather. He comes back with a chart shaped like a launch and a sentence he believes without reservation. The fund will return 27.5 per cent, guaranteed.

Within weeks the market turns and the fund falls. The chart is worthless, the confidence reads badly in hindsight, and the room that nodded along has quietly moved on to other things. The ordinary reading of this episode is that he was arrogant, or unlucky, or working with too little data — that with more variables, a longer sample and a better machine, the number would have held. That reading is comfortable, it is the one most institutions take, and it is wrong.

The model did not fail on its inputs. It failed on its grammar. Before a single figure was entered, the shape of the equation had already committed to two beliefs about the world: that a change in one quantity produces a change in another in fixed proportion, and that the people inside the system do not respond to what is being said about them. Both beliefs are false about finance. They are false in a particular way, too — not as small errors that cancel across many observations, but as errors that grow precisely when the stakes are highest.

A model that cannot represent a thing is never surprised by it. It is only confidently wrong about it.

This is the opening argument of Complexity-Informed Financial Regulation, and its real target is not the analyst but the rulebook. Supervisors write capital ratios, disclosure schedules and stress tests in the same grammar he used: a stable relationship between an input and an outcome, applied to institutions considered one at a time. When that grammar meets a system that loops, adapts and reorganises itself around whatever is being measured, the rules do not merely underperform. They can contribute to the instability they were drafted to prevent, and they can do it while every individual institution is in compliance.

What follows is the case in full. That financial systems behave like living ones, with flows, stocks, feedback and coupling that can be named precisely. That the mathematics of such systems is available and teachable and not especially exotic. That the behaviour which frightens regulators — bubbles, cascades, sudden evaporations of liquidity — is the ordinary output of ordinary rules interacting, rather than a moral failure by a handful of actors. And that regulation can be rebuilt to learn at the speed the system moves. None of it requires believing markets are unknowable. It requires giving up a specific kind of knowledge that was never actually there.

MACHINE OR ORGANISM

The machine metaphor has been doing quiet work in economics for a very long time, and it is worth saying what it buys before saying what it costs. A machine has separable parts. It can be diagrammed. A faulty component can be identified, removed and replaced without disturbing the rest, and the behaviour of the whole is the sum of the behaviours of the pieces. That is an enormously useful picture, and a great deal of financial supervision is built on it: examine each bank, certify each one sound, and the system composed of sound banks is itself sound.

The alternative picture is a forest. Trees convert sunlight into growth. Animals graze the undergrowth and carry seeds. Fungi break down what falls and return it to the soil. Nothing in the forest is in charge, and yet the forest is not chaos — it has structure, it has characteristic rhythms, and it recovers from most of what happens to it. Money in this picture is not a substance being counted but a nutrient in circulation, moving from savers through banks to businesses and back again as profit, wages and tax.

The parallel is more than decorative, and the book is careful about why. A metaphor earns its place by generating properties that can be checked, and this one generates five. Flows: capital, credit and information moving through the system at rates that can be measured. Stocks: the accumulations those flows build up — assets, debt, collateral, and also the softer inventories of regulation, convention and infrastructure. Feedback: the fact that what the system does changes what the system then does. Coupling: how tightly any one part is bound to the others. Emergence: the patterns that appear at the level of the whole and are present in none of the parts.

Add a sixth, which Nassim Nicholas Taleb named antifragility: the property of a system that does not merely survive shocks but is improved by them. A bone that remodels stronger after stress. A forest cleared of deadwood by fire and opened to new growth. The book is deliberate here and declines the sentimental version — nobody is being asked to welcome a crisis. The claim is narrower and more useful: some system designs convert disturbance into adaptation, most do not, and the difference is a matter of architecture rather than luck.

What this reframing changes is what a supervisor looks at. Under the machine picture, health is a property of components, and the natural instruments are capital buffers and institution-level examinations. Under the living picture, health is a property of the pattern — how concentrated the flows are, how tightly coupled the nodes, how much genuine diversity exists among strategies, how quickly information travels and to whom. Two systems can hold identical capital and have entirely different fragility, and only the second picture can see the difference.

This is not an argument for abandoning the balance sheet. It is an argument that the balance sheet describes a component in a network, and that a list of certified components has never been the same thing as a description of the network they constitute.

WHAT LINES ASSUME

It is worth being exact about what a linear model is, because the word gets used loosely and the precision is the whole point. In its simplest form it is the equation of a straight line: Y = mX + b. Y is the quantity of interest, X the quantity thought to drive it, m the rate at which one converts into the other, and b the value of Y when X is nothing at all. Increase the advertising budget by a tenth and sales rise by a tenth. The relationship is fixed, and it is the same at every value of X.

Look at what that structure quietly forbids. It forbids state dependence: the effect of a change cannot differ according to where the system already is, so a rate cut into a confident market and the same cut into a frightened one must produce the same result. It forbids memory: how the system arrived at its present position cannot matter. And it forbids reaction: the actors being described cannot notice the relationship and adjust their behaviour to it. Every one of these prohibitions is violated daily in finance, and not at the margins.

Take the relationship the entire industry is organised around, between risk and return. The linear version is a straight line — more of the first buys proportionally more of the second, indefinitely. The real relationship bends. Additional risk buys additional return up to a point, past which the risks overwhelm the compensation and expected return falls away sharply. A straight line cannot produce that shape. Fitted to data from the rising part of the curve, it will extrapolate confidently into the region where the curve has already turned over, and it will do so without any signal that something has changed.

The smallest honest counter-example in the book is the logistic equation, written dX/dt = rX(1 − X/K). Here X is the quantity growing, r is its intrinsic growth rate, and K is the carrying capacity — the ceiling the environment imposes. The first term, rX, is plain exponential growth, the arithmetic of anything that compounds. The whole argument sits in the second term. As X rises toward K, the factor (1 − X/K) shrinks toward zero and the growth rate falls with it. When X reaches K the expression is zero and growth stops.

A straight line has no way to notice how far it has already travelled. That is not a simplification of a market. It is a refusal of one.

That bracket is the system perceiving its own size and responding to it, expressed in eight characters. It is not advanced mathematics and it does not need a supercomputer. What it needs is a willingness to write down a model in which the rate of change depends on the current state, and the consequence of that willingness is enormous: the logistic equation produces an S-curve, rising fast, slowing, and settling, while the linear model produces a ray that never stops climbing. Given the same early data the two are nearly indistinguishable. Given enough time they disagree about everything, and the disagreement always arrives as a surprise to whoever chose the line.

Growth that notices its own ceiling looks like this from the outside: unhurried, and still moving.

THE LOOP

The book works the arithmetic rather than gesturing at it, which is the right decision, because the intuition only lands once the numbers have been moved by hand. Take a pool of investors in some instrument. Start with 100 of them, an intrinsic growth rate of 0.2 per year, and a carrying capacity of 500 — the total population plausibly reachable. Step forward in increments of a tenth of a year. At the first step the rate of change is 0.2 × 100 × (1 − 100/500), which is 16, so the population moves to 101.6. At the next step the rate is roughly 16.3 and the population reaches about 103.2.

Notice what happens as the iteration continues. The rate of change keeps rising for a while, because the pool is still small relative to its ceiling and there are many people left to recruit. Then it peaks, then it declines, and the curve flattens as it approaches 500. Nobody imposed the slowdown. No committee voted for it. The brake was in the equation from the beginning, in the term that compares the current size to the ceiling, and it engaged automatically as the ratio changed. That is negative feedback, and it is what a stable system looks like from the inside.

Now remove the brake, which is exactly what a speculative episode does. The book models a widely-discussed stock with an intrinsic growth rate of 0.5, a ceiling of $100 imposed by fundamentals and scepticism, and an opening price of $10. The first step adds $4.50, taking it to $14.50. The second adds about $6.20, to $20.70. The third adds about $8.21, to $28.91. Each increase is larger than the last. To anyone inside the episode this reads as confirmation: the move is accelerating, which is taken as evidence that the move is real, which brings in more buyers, which accelerates the move.

The mechanism has a name and no mystery to it. Rising prices attract buyers, buyers raise prices, and the loop is self-reinforcing until something interrupts it. The book sharpens this into a second equation in which the price at the next step is the current price multiplied by one plus a sentiment-driven rate of return, minus a correction term that grows with how far the price has drifted from its baseline: P(t+1) = P(t)(1 + r − s(P(t) − P₀)). The parameter s measures how strongly participants react to that drift — how quickly they start to feel the price is unreasonable.

Everything then depends on the relationship between r and s, and this is the finding worth carrying out of the chapter. If s is large relative to r, the correction bites early, the drift is limited, and the market oscillates around something defensible. If s is small relative to r — if enthusiasm is running strongly and scepticism has grown quiet — the negative term never becomes large enough to matter and the price runs away. Nobody has to be dishonest for this to happen. Two numbers describing a crowd's disposition are sufficient.

The cruel part is the timing. While the loop is running it is indistinguishable from a good investment thesis, and it is paying well, which makes it very difficult to argue with. The loop only becomes legible as a loop after it has reversed, at which point the same mechanism is operating in the other direction and the people describing it are the ones being carried by it.

THE ASYMMETRY

Here is a sequence small enough to hold in the hand. A technology firm announces a product. The stock, at $100, rises 20 per cent over a week to $120 as the news circulates and buyers arrive. Then a rumour surfaces that a competitor is preparing something similar. The rumour is unconfirmed and may be nothing. The stock falls 15 per cent in a single day, to $102. In price terms the round trip is nearly complete and almost nothing has happened. In every other respect the two moves are not remotely comparable.

The rise took a week and required a verified announcement. The fall took a day and required a rumour. That asymmetry in time and in evidentiary standard is the signature of nonlinearity, and it is not a quirk of this example. Upward moves are built by accumulation, as separate participants independently conclude that something is worth more. Downward moves are built by simultaneity, as participants conclude at the same moment that they would rather not be holding it, frequently because they can see each other deciding.

The same structure, run at national scale over several years, produced the crisis of 2008, and the book traces it channel by channel. House prices begin to rise, for reasons that are ordinary enough — cheap credit, real demand. Rising prices make existing owners wealthier, which supports spending, which supports the economy, which supports housing. Banks, watching collateral values climb, find their loans looking progressively safer and relax their standards. Lower standards admit more buyers. More buyers raise prices further. Every step in that chain is individually defensible and the chain as a whole is a loop.

Underneath it sits the nonlinearity that nobody priced. The relationship between house prices and mortgage risk is not proportional. As prices rise further above what incomes can service, the potential loss does not grow in step with the price — it grows faster, because the distance to fall increases at the same time as the cushion thins. Lenders reading a smooth upward series saw a stable business. They were looking at the flat early portion of a curve that was already bending, and the packaging of those loans into instruments sold onward meant the bend, when it arrived, was distributed to holders who had no view of the underlying at all.

Upward moves are voted for one at a time. Downward moves are voted for all at once.

Then rates rise, affordability erodes, prices plateau and begin to fall, and every channel reverses. Owners find they owe more than the house is worth. Foreclosures push more supply into a falling market, which pushes prices lower, which produces more owners underwater. Banks absorbing losses tighten standards, which removes the buyers who might have stabilised prices. The machinery that manufactured the boom does not break in the bust. It runs in the other direction at higher speed, and the shock that starts it can be modest, because its size was never what determined the outcome.

NOBODY DESIGNED IT

A colony of ants builds tunnels of real architectural sophistication — ventilated, structured, efficiently provisioned. No ant holds the design. Each follows a short list of local rules about food, pheromone and obstacle, and the nest is what those rules produce when a great many of them run near each other. A murmuration of starlings turns as a single body across an evening sky, and no bird is leading. Each is tracking its immediate neighbours, and the shape belongs to the flock rather than to any member of it.

This is emergence, and the book insists on the strict reading of the word. An emergent property is not merely a complicated outcome. It is a property that exists at the level of the whole and is present in no part, and which therefore cannot be found by examining the parts however carefully. No amount of studying one ant reveals the nest. No amount of studying one trader reveals the market.

The financial version runs like this. A hundred traders each hold a portfolio and each receives noisy signals — an analyst note, a headline, a conversation. One of them, reading a positive signal, buys. The price ticks up. Other traders observe the tick, and here is the hinge of the whole argument: they do not read it as one person's opinion. They read it as information, because in a market prices are the principal channel through which information is thought to travel. Some of them buy on that basis. The price rises further, which looks like more information, which brings more buyers.

The price is therefore both the output of the system and one of its inputs. Whatever else changes, that circularity does not, and it is enough on its own to generate booms, reversals and clustered volatility without anyone behaving irrationally. Each trader in that sequence acted sensibly on the evidence available. The evidence available was substantially manufactured by the sequence itself.

Price is the only evidence everyone can see, which is why a crowd watching prices is watching itself.

Two consequences follow, and both are uncomfortable for conventional supervision. The first is that the question asked after every crisis — who caused this — is frequently the wrong question, because the behaviour was produced by the interaction rather than authored by any participant. There may be misconduct alongside it, and misconduct should be pursued on its own terms, but removing the individuals does not remove the mechanism. The second is more hopeful. If pattern arises from local rules, then the rules governing local interaction are the effective lever, and those are precisely what a regulator can reach.

She is not waiting for certainty. She is deciding what she can hold through being wrong.

THE COUPLING

Draw the financial system as a network. Institutions are nodes. Loans, derivative exposures, repurchase agreements, custody relationships and shared clearing infrastructure are the edges between them. What the picture shows immediately is that the questions worth asking are not only about the size of the nodes. They are about the density of the edges, where they concentrate, and how quickly something travelling along them arrives somewhere else.

Coupling is the technical term for how tightly bound the parts are, and it is the variable most systematically under-measured. In a loosely coupled system a shock to one node is absorbed locally. Neighbours have slack, alternatives and time, and the disturbance dissipates before it propagates. In a tightly coupled system the shock passes straight through, because each node's obligations fall due on the same day as its claims, and there is no interval in which anyone can adjust.

The book's canonical illustration is the domino sequence of 2008, in which the failure of a small number of institutions propagated through interlocking exposures until the global system was at risk. The propagation did not require those institutions to be the largest in the world. It required them to be positioned where a great many edges met, holding obligations that a great many other balance sheets depended on. Position in the network, rather than size on the balance sheet, is what made them consequential.

There is a paradox buried in this that deserves to be stated plainly, because it inverts a piece of received wisdom. Diversification reduces risk for the individual holder, which is true and worth doing. But if every institution diversifies into the same broad set of assets by the same broadly standard method, their portfolios converge, and a system of well-diversified and nearly identical portfolios is less diverse than a system of concentrated and different ones. Every participant is then exposed to the same shock, and all of them respond to it in the same direction at the same time. Individually prudent behaviour has manufactured a collective fragility.

The instruments the book proposes follow directly from the picture. Map the network rather than only auditing its members. Identify the nodes whose failure would propagate furthest, and treat network position as a legitimate basis for requiring additional capital or additional diversification — not as punishment for size, but as pricing for structural consequence. Watch the concentration of exposures and the convergence of strategies as first-class indicators in their own right.

None of this replaces institution-level supervision, and the book does not claim it does. It sits beside it and answers a question the balance sheet was never constructed to answer: not whether this bank can survive its own losses, but what happens to everyone else on the day it cannot.

A HUNDRED TRADERS

If markets produce their behaviour through interaction, then a model that starts with aggregates has thrown away the mechanism before the first calculation. Agent-based modelling starts at the other end. Build the participants individually, give each its own rules, let them interact, and watch what the aggregate does. The aggregate is a result rather than an assumption, which is the entire methodological point.

The book builds one slowly enough to follow. Create a hundred agents. Give each two attributes: a risk tolerance between 0 and 1, and a current portfolio value. Let the price respond to the crowd's appetite, with the change in price equal to a sensitivity constant multiplied by the average risk tolerance across all agents. Set sensitivity at 0.1. If the average risk tolerance at this step is 0.6, the price change is 0.1 × 0.6, which is 0.06 — a rise of six per cent.

Then let the agents act on it. Each decides whether to buy, sell or hold according to a rule of its own. A threshold rule buys when the move exceeds some level and sells when it falls below another. A proportional rule adjusts holdings in step with the size of the move. And crucially, having gained or lost, each agent's risk tolerance shifts — success makes an agent bolder, loss makes it more cautious. Those revised tolerances feed into the next step's average, which sets the next price change. The loop closes with no external input required.

Richer versions give agents strategies rather than thresholds. A trend follower buys what is rising and sells what is falling. A value investor compares price against an estimate of intrinsic worth and acts on the gap. A noise trader responds to something closer to chance. The market clears through the aggregate of their orders, with the price change proportional to net buying pressure: ΔP(t) = α(ΣBuy − ΣSell), where α reflects how thin or deep the market is.

What comes out of running this is not a forecast and the book is firm about not pretending otherwise. What comes out is structural knowledge about how outcomes depend on composition. Raise the proportion of noise traders and volatility rises. Raise the proportion of trend followers and the simulation begins producing bubbles and crashes that nobody wrote into it. Thin the market — lower α — and the same order flow moves the price much further. These are findings about the shape of a system rather than predictions about a date.

Which makes the technique a laboratory for rules. A proposed circuit breaker, a transaction fee, a disclosure requirement, a margin change: each can be run against a population of agents and observed for the behaviour it induces, including the behaviour nobody intended. Farmer and Foley argued in Nature that the economy needs agent-based modelling, and the argument in this book is the operational version of that case. Policy that has been simulated against adapting participants before it becomes law is policy that has met at least one honest objection.

RULES THAT LEARN

Everything to this point converges on a single problem with the way financial rules are made. A rule is written at a moment, against a system observed at that moment, and takes effect on a system that has since adjusted to the rule. Participants read the requirement, understand what it measures, and reorganise their activity around it — not always cynically, often just by following the incentives as written. The rule then governs a configuration that no longer exists, and the activity it was meant to constrain has relocated to wherever the rule is silent.

The book's answer is not more rules or fewer, but rules built to update. Continuous monitoring comes first: standing measurement of volatility, of interconnectedness, of liquidity flows, sampled at the frequency the system actually moves rather than the frequency of the reporting calendar. A quarterly return describes a system that turns over in seconds. Early warning is impossible when the instruments run slower than the thing being watched.

Stress testing is the second instrument, and it needs reconstruction rather than extension. Conventional tests apply gradual, independent shocks — unemployment up by so much, prices down by so much — and ask whether each institution survives. That design assumes precisely the linearity this book spends thirteen chapters dismantling. A complexity-informed test asks different questions. What happens when several shocks arrive together and reinforce each other. What happens when every institution responds by selling the same asset. Where are the thresholds past which the system's behaviour changes character rather than merely degrading.

Third, diversity becomes a policy objective rather than an accident. If uniformity of strategy is a source of systemic fragility, then a supervisory framework that pushes every institution toward the same risk model, the same asset mix and the same response to the same signal has manufactured the danger it is meant to prevent. Encouraging a range of institutional sizes, structures and approaches is not indulgence toward the small. It is structural insurance, and it is bought by declining to standardise the things that do not need standardising.

Fourth, transparency is treated as infrastructure. In a system where behaviour emerges from what participants can see, what they can see is a design variable. Standardised reporting and open data on exposures allow the network to be mapped by more than one party, which means errors in the map can be found by someone other than its author. Information flow is not a courtesy owed to the public. It is a load-bearing component of how the system regulates itself.

A rule written against a fixed system is a photograph issued as a map.

The disposition underneath all four is best put as gardening rather than engineering. A gardener does not specify each branch. A gardener conditions the soil, manages the light, removes what is crowding, and lets the growing happen — intervening in the conditions that shape outcomes rather than dictating the outcomes themselves. Applied to finance this means regulating coupling, concentration, information and diversity, and accepting that the specific path the system takes through those conditions is not knowable in advance and never was.

WHAT IT ASKS

The argument narrows to something an individual can act on, and the first move is a change of goal. Stop trying to predict and start trying to adapt. These are not two ways of saying the same thing. Prediction asks what will happen and commits accordingly, which means a wrong answer is expensive and being right is the only route to surviving. Adaptation asks what range of things could happen and builds a position that does not fail badly in any of them. The second goal is achievable. The first, in a system that reorganises around its own forecasts, is not.

Diversification follows, and it means more than holding several things. Spread across asset classes, across geographies and across industries, because the point is not to own many positions but to own positions that respond differently to the same event. Two holdings that fall together in every crisis are one holding with extra paperwork. Rebalance on a schedule rather than on a feeling: as values drift, the portfolio's actual risk drifts away from the intended risk, and periodic rebalancing restores the shape without requiring anyone to have a view.

Hold a long horizon, and hold it for a reason rather than as a platitude. Short-term price moves are the emergent output of millions of interacting decisions, which is precisely the category of behaviour the book has shown to be structurally unpredictable. Attempting to trade them is attempting to forecast an emergent property from the outside. The longer horizon is where the fundamentals that agent-based models treat as anchors actually assert themselves.

And then antifragility, which asks something harder. Taleb's insight is that shocks are not uniformly destructive — some systems, and some positions, are improved by disturbance. In a portfolio this means keeping reserves that turn a fall into an opportunity to buy rather than a forced sale. In an institution it means designing so that small failures are survivable and informative, because a system that never fails small is one that has quietly concentrated all of its failure into a single event it cannot survive at all.

The purpose of a reserve is not to feel safe. It is to be the buyer on the day everyone else is selling.

What Complexity-Informed Financial Regulation ultimately offers is not a new set of forecasts but a change in what counts as understanding. The thirteen chapters move through the limits of linearity, the anatomy of complex adaptive systems, emergence and self-organisation, feedback and nonlinearity, and agent-based simulation — each chapter opening with a story, working its mathematics out by hand with numbers a reader can follow, testing it against market behaviour, and turning it into practice for institutions and individuals alike. The market was never a machine with a manual somebody had mislaid. It is a living arrangement, and the useful question was never how to predict it. It is how to build something that keeps standing while it moves.

Free to read

Free to read, and free to hear. Every chapter of every book in this house, and every narration of it, is open to anybody. No account, no card, nothing to cancel.

What is in it

• Introduction: The Limits of Linearity in Financeopen this in the house search
• Financial Systems as Complex Adaptive Systemsopen this in the house search
• Emergence and Self-Organization: Understanding Market Dynamicsopen this in the house search
• Feedback Loops and Nonlinearity: Drivers of Financial Instabilityopen this in the house search
• Agent-Based Modeling: Simulating the Behavior of Financial Marketsopen this in the house search

A straight line cannot bend. That is not a simplification of a market; it is a refusal of one.

Every crash is a feedback loop that was called a trend while it was still paying.

Diversity is not a courtesy in a portfolio. It is the only insurance a coupled system sells.

Nobody authored the pattern. Everybody supplied a rule, and the rules met.

Keep looking

Every phrase on this page opens into the house search. The shelf holds Living Systems Economics and six other shelves, and the reading is free.