Haute Lumière · The reading

Nothing happened, and the price moved anyway

The case for modelling a market as a crowd of unlike people rather than as one rational average of them.

She has read the same paragraph four times. The argument is simple, and the consequences of believing it are not.

THE MISSING PERSON

The book opens with a finance professor in tweed, mid-lecture on the Efficient Market Hypothesis, while a company that sells inflatable flamingos for backyard pools goes vertical on the ticker behind him. The financials are mediocre. There is no breakthrough technology. There is internet enthusiasm and a dancing mascot. His students, who have been following the whole thing all week with some amusement, understand it better than he does, and not because they know more finance. They know more people.

It is a comic scene doing serious work, because the professor's apparatus is not wrong in its own terms. Black-Scholes prices an option cleanly, and the book walks the arithmetic rather than gesturing at it: a stock at $50, a strike at $55, a year to run, a risk-free rate of 5 percent, volatility at 20 percent. That gives d1 of about −0.196 and d2 of about −0.396, cumulative normal values of roughly 0.423 and 0.348, and a call worth about $3.07. The arithmetic is sound. The question the book asks is where the ingredients came from.

Underneath the formula sits a single sentence about the world: dS/S = μdt + σdZ. A tendency, and a tremor. The tendency is expected return. The tremor is a random draw from a normal distribution, scaled by a number called volatility that you hand the model before you start. Read it as a description of a trading floor and notice what it contains: nobody panics, nobody copies their neighbour, nobody is forced to sell at the worst possible hour because a redemption request landed that morning. The turbulence is an input, not a consequence.

The tremor in the standard equation is a number you supply. In an agent model it is something the crowd does to itself.

The regression version of the same habit is the cheaper confession. Stock Price Change = β₁ × Interest Rate Change + ε, with β₁ a constant sensitivity and ε a residual absorbing everything the model preferred not to represent. Crowding lives in ε. Fear lives in ε. The fact that the seller and the buyer are two different kinds of person with two different time horizons lives in ε. When one of those things dominates a week, the equation does not break loudly; it simply reports a large residual and carries on, which is a far more expensive kind of failure than an error message.

So the argument is not that the classical models are false. It is that they are models of an average, and an average has never placed a trade. What is missing is not resolution, and not more data. It is persons, plural, unalike, each acting on partial information in their own interest, and the whole apparatus of consequence that follows from their being different from one another.

THE FLOCK

Every agent-based argument rests on one observed fact, and the book states it with birds. A bird in a flock follows about three rules: stay near your neighbours, avoid collisions, react to the predator. Nothing in those rules mentions the shape of the flock, and no bird is holding the shape in mind. The shape appears anyway, shifts, splits, reforms, and is impossible to derive by studying one bird carefully. The book's word for this is emergence, and it is the whole hinge of the method.

The same phenomenon is set out at a farmer's market, which is closer to finance than the birds are. A woman puts back a peach because it is not ripe enough for this afternoon. The farmer hears the real requirement inside the complaint and hands her a plum. A sale happens. Multiply that interaction by a few thousand and a price level exists, stalls thrive or fail, and tomorrow's supply changes in response, though nobody set the price and nobody planned the market's behaviour.

The move from that scene to a stock exchange is smaller than it looks. Each participant has a rule, some money, a tolerance for being wrong, and a partial view. What the exchange produces at the end of a day is not a summary of those rules. It is a different kind of object altogether: a price path, a volume profile, a distribution of returns with fat tails and clustered volatility that nobody wrote down as a target.

No trader intends a bubble. A bubble is what a room full of reasonable people produces when each of them is watching the others.

This matters most for the events that classical finance handles worst. A crash is not usually a mistake committed by one large actor. It is a configuration: a price fall that trips a stop, which becomes a sell, which deepens the fall, which trips the next stop. Every individual decision in that chain can be defended on its own terms, and the aggregate is a catastrophe. There is no contradiction there, only a level confusion. The rationality lives at the level of the agent; the pathology lives at the level of the system.

Which is why the book insists on a particular discipline. You do not model the crash. You model the traders, you give them their rules, you let them run, and you watch to see whether a crash appears without being asked for. If it does, you have learned something about the mechanism. If it never does, no matter how you shake the parameters, you have learned something too, and it is the more interesting finding of the two.

WHAT AN AGENT IS

An agent, stripped to its parts, is four things: a state, a rule, a budget, and an information set. The state is what it holds. The rule is what it does when it looks at the market. The budget is what it can afford to do. The information set is what it is allowed to see, and this last one is the part most models get wrong by being too generous.

The book's simplest population has one hundred agents and one asset priced at $10. Each agent has a risk tolerance on a scale, an order size that reflects both capital and nerve, and a price at which it will buy or sell. Some read only the price chart. Some have access to research the others do not. That asymmetry is not a defect in the simulation. It is the thing being simulated, because information asymmetry is where trading profit actually comes from.

Preferences get a formal shape too, and the one the book uses is the power utility function, U(x) = x^(1−ρ)/(1−ρ), where ρ is how much the agent dislikes uncertainty. At $10,000 of wealth and ρ of 0.5, utility works out to about 141.42. The number is meaningless on its own and that is exactly the point: it only matters in comparison. Would this agent rather hold the asset, or the cash, given what it believes about tomorrow? Utility turns a personality into a comparison the machine can run a million times.

Heterogeneity is not noise added for realism. It is the engine. Identical agents produce a market where nothing trades.

The book returns to that last point repeatedly, and it is worth sitting with. If every agent has the same information, the same beliefs and the same risk appetite, there is no trade, because for any proposed transaction both sides agree on the value and neither has a reason to move. Trade requires disagreement. Volume is disagreement made visible, and liquidity is disagreement standing ready. A model that averages the disagreement away has deleted the reason markets exist.

So the agent classes in the book's simulations are deliberately uneven. Value investors work from fundamentals and are slow. Momentum traders work from recent price and are fast. Noise traders act on sentiment, headlines and something less defensible than either, and they are not there as a joke: without them the price discovery process has no grit in it, and the simulations behave more tidily than anything ever observed. The market maker sits across from all of them, quoting both sides, and is the only participant whose job is the system rather than a position in it.

The work goes faster on paper, because paper does not offer to help.

THE LOOP THAT FEEDS

Once agents can see the price, and the price is made by agents, the system has a loop in it, and loops are where the interesting behaviour lives. The book names the two kinds plainly. A positive loop amplifies: a price rises, the rise attracts buyers who read rises as signals, and their buying raises the price further. A negative loop damps: a price falls far enough that bargain hunters step in, and their buying slows the fall. Both are present in every market at all times, and which one dominates is not fixed.

The cleanest demonstration in the book uses two agent types, fundamentalists and chartists, and lets them share one stock. Sixty fundamentalists value the company from its expected earnings. Forty chartists read the last few days of price. Good news arrives: earnings look five dollars better than expected. The fundamentalists revise their target upward and buy, which moves the price, which the chartists read as momentum, so they buy too, which moves it further. The stock begins at $100 and reaches $108 — and only part of that move came from the news.

The other part came from the chartists reading the fundamentalists' reaction and treating it as independent evidence. This is the mechanism behind overshoot, and it is why a stock so often travels further on good news than any sober valuation supports. Nobody in the chain is irrational. The chartists are correctly identifying momentum. The momentum is correct. It is just not information about the company.

A price that is read as a signal stops being only a price. It becomes an input to the process that produces it.

The reversal has the same shape running backwards. A competitor announces something, fundamentalists mark their targets down and sell, and the chartists — who bought the rise on momentum — now see the opposite pattern and sell into it. What in a classical framework would be a modest downward revision becomes a slide, because two populations with different reasons are pushing the same direction at the same time. The book is careful that this is not presented as a prediction. It is a mechanism, and its value is that you can watch it run and vary it.

And that is what varying the parameters is for. Raise the proportion of chartists and the simulated market gets more volatile and more prone to sustained trends. Raise the fundamentalists and it becomes calmer and slower to respond. Change belief strength — how stubbornly each agent holds its view against contrary price action — and the character of the market changes again. None of these knobs exists in a representative-agent model, because a representative-agent model has only one kind of person to adjust.

THE BOOK OF ORDERS

Beneath the price there is a stream of buy and sell orders arriving, resting, cancelling and executing, and the book treats this as the actual substance of the market. The chapter opens with a trader who has built a position all week on a rumour, sees a million-share buy order hit the tape, and watches the price fail to move. Sell orders arrive alongside the buys and absorb them. His conclusion, and the chapter's, is that the rumour was never the mechanism. The order book was.

The modelling is deliberately simple so the moving parts stay visible. Orders arrive as a Poisson process at rate λ — in the worked example, five per second, which means the probability of a quiet tenth of a second is e^(−0.5), about 0.607. Each arriving order is a buy with probability 0.6 and a sell otherwise. Size is drawn from an exponential distribution, f(x) = αe^(−αx), with α of 0.01 giving an average order of a hundred shares. Three draws, and you have a simulated order: direction, size, timing.

What makes this worth doing is what happens when the orders meet. A market with five hundred shares wanted and three hundred offered has an imbalance, and the price moves to find the rest — in the book's example, from $10.00 to $10.25. That quarter is not a judgement about the company. It is the cost of demanding immediacy from a book that did not have enough resting on the other side, which is the working definition of illiquidity.

Liquidity is not a property of an asset. It is a property of who happens to be willing at the moment you ask.

This is the correction the chapter is really making. Classical finance treats volatility as a fixed characteristic of a security, estimated from history and carried forward. Here it is an outcome of the book's depth, which changes minute to minute with who is present. The same news landing on a thick book moves the price a little; landing on a thin one it moves it a great deal. Volatility clusters because liquidity clusters, and liquidity clusters because participants watch each other and withdraw together.

The practical consequence in the book is unglamorous and immediately usable. A market order demands immediacy and pays whatever the book charges for it at that instant. A limit order names a price and waits, giving up certainty of execution to avoid paying that toll. In a calm book the difference is small. In a disturbed one it is the whole result, and the disturbed moments are exactly when the impulse to use a market order is strongest.

There is a second consequence for anyone building these models. Order flow is the only place where individual behaviour and aggregate price meet mechanically, with no averaging step in between. Model the flow and the price is produced. Model the price directly and you have assumed the answer.

THE TUNING

A simulation with invented numbers in it is a story with equations attached, and the book is blunt about this. Calibration is the step that decides whether the thing is an instrument or an illustration, and it gets a full chapter with the arithmetic shown. The image offered is tuning strings: the instrument is built, and until the tuning is done, what it makes is not music.

Two routes are set out. Maximum likelihood asks which parameter values make the observed data most probable — you write the likelihood of the data given a mean and a standard deviation, and search for the pair that maximises it, starting from the sample mean and sample deviation and letting an optimiser do the rest. Bayesian inference does something different: it starts from a prior. If you already believe this stock's annual volatility sits somewhere between 10 and 20 percent, you encode that as a distribution centred near 0.15, combine it with the likelihood, and read the posterior.

The difference between the two is not technical fussiness. Maximum likelihood declares that you knew nothing before the data arrived, which for anyone who has watched a market for a decade is a false statement dressed as modesty. A prior makes the assumption explicit and therefore arguable. Someone can look at your prior and say it is too tight, or centred wrong, and that conversation is a feature.

A model with no stated prior does not lack assumptions. It has hidden them in the choice of estimator.

Then there is validation, which the book separates carefully from calibration and which is where most of the discipline sits. Fitting parameters to history is easy, and a model with enough knobs will reproduce any past you show it. The test is out-of-sample: hold back data the model never saw during fitting, run it forward, and compare. Not the exact path — no one expects that — but the statistical texture. Does it produce the right amount of volatility? Does volatility cluster the way real volatility clusters? Are the tails the right weight?

The book also names the failure mode that no amount of care removes. Several different parameter sets will often fit the data equally well, which means the data alone cannot tell you which mechanism is operating. That is an honest limit, and the response it recommends is not to pick the prettiest one. It is to carry the candidates forward together and see where they disagree about something you can still go and check.

Data itself gets its own treatment: prices, volumes, order-level records where you can get them, sentiment, macroeconomic series. And one further source the book takes seriously — expert elicitation, which is to say asking practitioners how they actually decide. That is qualitative, unfashionable, and frequently the only route to a parameter that no dataset contains.

Nothing here is urgent. Thinking that holds under pressure is mostly done before the pressure arrives.

WHAT IT WILL NOT DO

The book says no, repeatedly, and says it in the chapters where saying yes would have been easiest. Agent-based models are not crystal balls. They will not tell you Friday's close. Anyone selling that has either misunderstood the method or is counting on you to.

The reason is structural rather than a matter of insufficient effort. These systems are sensitive to initial conditions and to the sequence in which things happen. Two runs with identical parameters and a different random seed produce different price paths. That is not a bug being apologised for; it is the model correctly reporting that the world it represents is one where the same setup can go several ways. A method that returned one confident path would be lying about the phenomenon.

The method trades a precise answer you cannot rely on for an approximate understanding you can.

What it gives instead is the distribution and the mechanism. Run it ten thousand times and you learn what range of outcomes this configuration produces, how often the bad tail appears, and which parameter makes the bad tail fatter. That last one is the genuinely valuable output. Knowing that a crash becomes four times more likely when a particular class of agent exceeds a certain share of volume is worth considerably more than a point forecast, because it is actionable in advance and it does not expire on Friday.

The second honest limit is that a model can only exhibit behaviour its rules permit. If you did not give any agent a reason to withdraw liquidity under stress, your simulated market will never have a liquidity crisis, and its silence on the matter means nothing. This is the trap: absence of a behaviour in output is evidence only about your rule set, and it is very easy to read as evidence about the world.

The third is overfitting, and it is the most seductive because it feels like success. Add enough agent types and enough parameters and you will match history to a fine degree, and you will have built an elaborate description of a past that is not coming back. The book's guard is the same one good empirical work has always used: fewer moving parts than you are tempted by, every parameter justified by something outside the fit, and out-of-sample testing treated as a pass or fail rather than a formality.

Read straight, these limits are the most persuasive part of the case. A method that says plainly what it cannot do is describing something real. The book's own summary of the situation is that the aim is not to predict the future but to understand the mechanisms that generate it, and it holds that line even where a bolder claim would have sold better.

BUILDING ONE

The procedure is given as a sequence, and it is the same sequence whether the target is a bank's risk desk or a household's retirement account. Define the scope. Identify the agents. Write the interaction rules. Choose the platform. Calibrate and validate. Experiment. Refine. The book's only warning about the order is that skipping the first step is the most common way to waste a month.

Scope means naming the specific question. Not "model the market" — that is not a question and has no answer. Something closer to: what happens to this portfolio's drawdown if the share of momentum-following capital doubles? A question with a shape tells you which agents belong in the model and, more usefully, which do not. Every agent you include costs you a calibration problem, so the right number is the smallest one that can produce the behaviour you are studying.

Identifying agents means looking at the real participants and grouping them by behaviour rather than by name. Retail investors reacting to headlines and social sentiment. Institutions working from technical and fundamental signals with slower clocks and larger size. Market makers quoting both sides and managing inventory. Algorithms, which are agents with no patience and perfect discipline. The groups that matter are the ones whose rules differ from each other in ways that show up in the price.

The right number of agent types is the smallest number that still produces the behaviour you are trying to explain.

The rules are where judgement enters, and the book keeps them modest on purpose. A retail agent buys when a name is trending. An institutional agent sells when a valuation ratio passes a threshold. A market maker widens its quotes as order flow gets one-sided. Each rule should be simple enough to state in a sentence and to defend to someone who trades for a living. Complexity in the rules migrates directly into the calibration problem, and the trade is almost never worth it.

For the platform the book names the working options: NetLogo, which is the gentlest entry and genuinely usable without much programming; Mesa, MASON and Repast for models that need to scale. All are open source, all are established in the literature, and the honest note is that none of them removes the hard part, which was never the code.

Then the loop, which is the part that separates a model from a toy: calibrate to history, validate out of sample, run scenarios, and refine. Scenario work is where the value shows up — rates rise, a regulation lands, sentiment turns, a large participant withdraws. Each scenario is a run, not a guess, and the output is a distribution of consequences rather than an opinion. And the loop does not close. Markets change composition, the rules that fitted last year describe a population that has moved on, and a model that is not recalibrated is a historical document being consulted as if it were current.

THE LIVING SYSTEM

The frame the book returns to in every chapter is biological, and it is not decoration. A coral reef is not a pile of corals. It is polyps following local rules — grow toward light, retreat from danger — producing between them a structure none of them contains, with predators, competition and dependencies running through it. The book's claim is that a market has the same architecture, and that the vocabulary developed for living systems transfers with unusual precision.

Four terms carry most of the weight. Stocks are what accumulates: capital invested, shares outstanding, confidence held. Flows are the movements that change them: purchases, redemptions, issuance, the arrival of information. Feedback is how a flow alters the thing that produced it. Coupling is how a change in one part reaches a part that looked unrelated — a housing market and a bank and a pension fund, connected through exposures nobody had drawn on a single page. Most systemic risk is a coupling that existed before anyone described it.

The fifth term is the one the book treats most carefully, because it is the easiest to misuse. Antifragility, in Taleb's sense, is the property of a system that does not merely survive stress but is improved by it: the forest that needs the fire to clear its deadwood. The book notes that markets sometimes display this, adapting after a crisis into a more robust arrangement than they had before, and it does not dress the claim up further than that. Antifragility is a property some systems have under some conditions, not a reassurance available in advance.

A crash is not a failure of the market's logic. It is the market's logic, running at a level nobody was watching.

Where this lands for a reader without a research budget is more concrete than it sounds. It changes what a portfolio looks like. Diversification, read through coupling, is not owning many things; it is owning things whose holders behave differently under stress, because assets held by the same people in the same funds with the same redemption terms are one asset wearing several names. That distinction only appears when you look at the participants rather than the instruments, which is the method's central move applied to a personal balance sheet.

It also changes what a bad week means. If prices are produced by a crowd with feedback in it, then a fall that outruns any news is not a mystery and not a verdict on your judgement. It is a loop completing. Knowing the mechanism will not make the number smaller, but it does answer the question that makes people sell at the bottom, which is whether something is happening that they have failed to understand.

Farmer and Foley made the case in Nature in 2009 that the economy needs agent-based modelling. The intervening years have made the argument less controversial and the tooling considerably better, and this book is an attempt to hand the method over intact: the reasoning, the mathematics worked rather than gestured at, the calibration, and a clear account of where it stops. The reading is free, all thirteen chapters of it. What is for sale is a copy to keep.

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 Traditional Financeopen this in the house search
• Agents and Interactions: Building Blocks of Financial Marketsopen this in the house search
• From Micro to Macro: Emergence in Agent-Based Modelsopen this in the house search
• Calibrating the Market: Data and Parameter Estimationopen this in the house search
• Simulating Order Flow: Liquidity, Volatility, and Price Dynamicsopen this in the house search

A model that assumes calm cannot explain a panic. It can only name it and carry on.

The representative agent is an average of people, and an average of people has never placed a trade.

Explanation is what the method sells. Prediction is what people came for and what it declines to offer.

Identical agents produce a market in which nothing trades. Disagreement is not friction; it is the mechanism.

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.