Haute Lumière · The reading

The economy is not a machine, and that is why the forecasts keep missing

What changes when you stop asking who is steering and start asking what the system is doing to itself.

Between the fourth run that failed and the one that held. What changed was a parameter, not an idea.

THE FORECAST PROBLEM

The forecasts are revised every quarter, and the revision is rarely small. This is not a story about careless economists. It is a story about a mismatch between an instrument and the thing it is pointed at. A forecast is built for a machine — something with inputs, a transfer function, and outputs that follow from them after a lag you can estimate. Push the interest rate here and demand moves there, by roughly this much, in roughly that many months. The arithmetic inside such a model is usually impeccable; what fails is the assumption underneath it, that an economy is the kind of object that has a transfer function at all.

An economy has no outside. Every actor inside it is also a forecaster, adjusting to what everyone else is expected to do, including adjusting to the forecast itself. The moment a model is published it becomes an input to the system it describes. Weather holds no opinion about the forecast; markets do. That single difference dissolves the clean separation between observer and observed which makes prediction work in physics, and it is why a model fitted beautifully to the last twenty years of data can be useless about the next two. The data was produced by a system that has since changed, partly in response to having been modelled.

A weather system does not read the forecast. An economy reads it, and then it is a different economy.

The alternative is not to abandon rigour. It is to change what rigour is aimed at. Complex-systems economics gives up point prediction and takes up something harder to sell and more useful to hold, which is structure. Which loops in this system amplify and which damp. Where the network is dense enough that trouble travels. Which quantities are stocks that take a decade to rebuild and which are flows that can stop on a Friday. A practitioner who knows those three things cannot tell anyone the level of an index next June. They can say which failures are possible, which are cheap, and which cannot be recovered from — and that is the information a decision actually needs.

This is the argument the book makes, chapter by chapter, and it is worth stating in its hardest form before the machinery arrives. Economies are not badly predicted because the models are crude. They are badly predicted because prediction is the wrong ambition for a system whose most consequential behaviours — booms, crashes, clusters, whole industries appearing where none existed — are things it does to itself rather than things done to it.

ORDER WITHOUT A DESIGNER

The book opens in Naples with a man called Salvatore, who makes good pasta by hand in a shop nobody walks past. A photographer called Sofia comes in lost and hungry, eats, and starts documenting the place — the process, the customers, the decor. She posts the pictures. Slowly, tourists start turning off the main streets, locals come to see what the noise is about, and within a season Salvatore has hired help and opened a second location. The tempting reading is that this is a story about marketing. The book's reading is stricter, and it is the foundation of everything after it.

Nobody designed that outcome. Salvatore did not plan a photographer. Sofia did not plan a business. The people who eventually queued were responding to other people's responses, and no participant held the shape of the thing they were building. This is emergence: a pattern at the level of the whole that no member intended and none could have produced alone. The same structure runs a farmers market, where each vendor responds only to local conditions — what is ripe, what is selling, what the next stall is charging — and none of them holds a picture of the market entire. Prices converge anyway. Gluts clear. Gaps fill.

Self-organisation is the mechanism underneath. Decentralised actors, local rules, no conductor, and structure appearing regardless. The book's point is that this structure is not a metaphor; it does work. It allocates scarce things among people who never meet. And it explains why intervening at the level of an outcome so often fails: the pattern is not stored anywhere that can be edited. It is continuously recreated by the interactions that produce it, so changing the rules or the connections changes the pattern, while changing the pattern alone lets it re-form by morning.

The price was not decided by anybody. It was produced by everybody, and it is written down in no one's head.

There is a corollary here that cuts against the romance, and the book does not skip it. Self-organisation is not a synonym for a good outcome. Emergent patterns include speculative bubbles, segregated neighbourhoods, cartels, and the quiet concentration of an industry into three hands. The same mechanism that clears the market produces the crash. Complexity is a description of how order arises, never a promise that the order is one anybody would have chosen — and reading emergence as a guarantee of efficiency is the most common way the idea gets misused by people who have only met half of it.

STOCKS AND FLOWS

A flow is a rate: money moving, goods shipped, people hired, an idea passed from one person to the next. A stock is an accumulation: factories, trained workers, roads, reserves, installed knowledge, trust. Flows are measured per unit of time; stocks are measured at an instant. Almost all public argument about the economy is conducted in flows — this quarter's growth, this month's hiring, last week's orders — because flows are what get published on a schedule. The distinction sounds like accounting pedantry until the day it decides what is possible.

Stocks set the ceiling. A country can raise its investment flow overnight and still wait fifteen years for the stock of engineers that investment was meant to use. A firm can double its marketing spend this month and discover that its stock of reputation moves on a different clock entirely. Stocks are also what makes a shock permanent rather than temporary: a flow that stops can restart in a week, while a stock that has been destroyed is rebuilt only at whatever rate the flows allow. That asymmetry is the difference between a bad year and a lost decade, and it is invisible in any account that watches only the monthly numbers.

The book writes the simplest version of the interaction out in full. Growth in a limited environment follows dX/dt = rX(1 − X/K): change proportional to what already exists, damped by how close it has come to what the environment can carry. With the book's rabbits — an intrinsic growth rate of 0.2, a carrying capacity of 100, a starting population of 10 — the first year adds 1.8 animals, and each year adds less as the population nears the ceiling. The same equation applied to firms with r = 0.2 and K = 50 takes ten firms to about twenty-six after five years. Nothing about the firm changed. The room changed.

A flow that stops can start again in a week. A stock that is destroyed is rebuilt at whatever rate the flows allow.

What makes this a living-systems idea rather than arithmetic is that K is not a constant. Carrying capacity in an economy is itself built out of stocks — infrastructure, institutions, credit, skills, attention — and those stocks are laid down by the very flows that are pressing against the ceiling. So the ceiling moves, and it moves in response to the growth testing it. A new technology raises it. A financial crisis lowers it, sometimes for years. A model that fixes K produces a tidy S-curve; an economy produces an S-curve whose ceiling is being renegotiated by the people climbing toward it.

The diagram under her hand has eleven nodes in it and already behaves in ways she did not put there.

THE LOOPS

There are two kinds and the naming trips almost everyone. A balancing loop counteracts: prices rise, demand falls, prices settle, the way a thermostat answers a cold room. A reinforcing loop amplifies: unemployment rises, spending falls, demand falls, firms cut staff, unemployment rises further. The book is careful about the vocabulary because the technical word for the second kind is positive, and positive here means self-amplifying rather than desirable. The most powerful reinforcing loop in the whole book is a depression feeding on itself.

The mathematics can be small and still say something. The book sketches an economy in two lines: dC/dt = rC + αI, and dI/dt = βC − δI. Consumer spending grows on itself and on investment; investment grows on consumer spending and decays through depreciation. Four parameters, no politics, nothing but a statement about how two quantities are coupled. Setting both rates to zero locates the equilibria, and the interesting discovery is that depending on those four numbers a system of this shape can have one stable point, several, or none at all.

That last case is where the argument earns its keep. A system with no stable equilibrium does not sit quietly waiting for something to hit it. It generates its own turbulence out of entirely ordinary internal dynamics. Which means an explanation of a downturn that requires an external cause may be searching for something that was never there — the cause can be the arrangement. That is a materially different diagnosis from a shock arriving from outside, and it implies a different repair: alter the loop, not merely the weather.

Positive does not mean good. It means the system is now an argument the system is having with itself.

Delay is the third ingredient and the one most often left out of the sketch. A loop with a lag overshoots, because by the time the correction lands, the thing being corrected has already moved past the target, and the correction becomes the next excess. Housing supply, capital investment, training a workforce and planting an orchard all carry lags measured in years, and all of them oscillate for exactly that reason. Cycles require no cyclical cause. A balancing loop with a long enough delay will manufacture them unaided, forever, out of nothing but its own good intentions.

THE NETWORK

The book spells out the vocabulary rather than assuming it. The degree of a node is how many connections it has. Path length is the shortest number of hops between two nodes. Density is actual links over possible links — five nodes can support ten links, and if only four exist the density is 0.4. Centrality measures how much of the traffic must pass through a given node. Clustering asks whether nodes gather into groups that talk more among themselves than outside. Five plain measurements, each computable from a table of who deals with whom.

None of these is decoration. Each is a risk measure wearing a neutral name. A dense network with short paths moves information quickly, moves capital quickly, and moves failure quickly — the speed is one property, not three. A high-centrality node is an efficiency in fair weather and a single point of failure in foul. The small-world result the book cites, from Watts and Strogatz, shows that a handful of long-range links collapses path length across an entire network, which is excellent for the diffusion of a good idea and identical in mechanism to the propagation of a bad one.

The book then does something useful with it and builds a portfolio as a network. Three assets, correlations measured rather than assumed: a tech stock and a government bond at −0.3, the tech stock and a property trust at 0.6, the bond and the trust at 0.2. Drawn as three circles joined by lines whose thickness is the strength of the relationship, diversification stops being a list of names and becomes a shape you can look at. Owning many things is not diversification when the many things are tightly coupled to each other.

The same short paths that carry a good idea across a continent in a week carry a failure across it in a day.

And correlations are not constants. They describe the current state of a network that rewires under pressure, and the rewiring tends to run one way: connections tighten in a crisis. A portfolio that looked pleasantly spread out in calm conditions can converge toward a single underlying factor at precisely the moment spread was the point. This is why a network view has to be maintained rather than computed once and filed. Interdependence is simultaneously the asset and the exposure, and which of the two it is on any given morning depends on the weather.

WHY CRASHES HAPPEN

Nonlinearity is the property that makes a system stop being polite. The book's image is a pendulum: pushed gently it traces a predictable arc, and pushed hard enough it stops tracing anything you can write down. Small changes stop producing small effects. The same input, applied to the same system in two different states, produces two unrelated outcomes. Every intuition built on proportionality — twice the stimulus, twice the response — fails here, and most of the intuitions people carry about economies were built on exactly that.

The worked case in the book is small and honest about its own size. A tech stock at $100, a hundred shares, ten thousand dollars committed. The price rises five per cent a week for three weeks — $105, then $110.25, then $115.76 — and the position is worth $11,576. Then the company announces a delay to its product, and the price gives back three per cent a week for two weeks, to $112.19 and then $108.87, leaving $10,887. That is a loss of about $689 from the peak. Almost nothing about the company's long-run value moved in five weeks. What moved was the direction of a loop.

A bubble is a reinforcing loop in which price has become the evidence for price. Buying lifts the number; the lift is read as information; the information attracts buying; the cycle needs nothing to be true about the underlying asset in order to run. It continues while the flow of new participants holds, and it reverses through the identical mechanism when that flow thins, because a falling number is also read as information. The rise and the collapse are not two phenomena with two explanations. They are one loop, running in each direction.

A bubble is not a mistake about value. It is a loop in which the price has become the evidence for the price.

The word the book uses for the moment of turn is bifurcation: a threshold at which a system's behaviour changes character rather than degree. This is why warning signs are so weak in advance. Through nearly the whole run-up the system genuinely is stable, and it reports itself as stable, right until a parameter crosses a line that nobody marked. The practical consequence is a change of posture rather than a better alarm. The useful question is not when it will break but what state it is currently in, and how far that state sits from the boundary.

She has stopped asking what the number will be and started asking what would have to be true for it.

AGENTS, NOT AVERAGES

Aggregate models replace a population with a representative actor and reason about the average. It is convenient and it is occasionally catastrophic, because the average tells you nothing about the distribution, and in a system full of thresholds and networks the distribution is the entire story. Average household savings can look healthy while a third of households cannot absorb one missed paycheck. The average is not wrong; it is answering a question nobody asked, and it hides precisely the tail that decides whether a policy works.

Agent-based modelling builds the population instead. The book's diffusion example gives each entrepreneur two attributes — whether they have adopted a new technology, and the list of people they talk to — and makes the probability of adopting a logistic function of exposure to adopters plus an intrinsic propensity. With the book's parameters, β at 2 and γ at −1, an entrepreneur connected to two others of whom one has adopted carries roughly a 0.73 chance of adopting in that step. Run every agent, repeat, and an adoption curve appears in the aggregate that was written into no agent's rules.

The market version puts heterogeneous risk appetites into a single stock. Some agents hold a low tolerance and buy when the price is soft, some hold a high one and buy on expected growth, each submits orders, the book matches them, and the price for the period falls out of the balance between buying and selling. A move from $50 to $55 pays each agent in proportion to what they were willing to carry. Volatility, clustering, bubbles and crashes then appear at the level of the market without anyone having written volatility into a rule.

No agent in the model contains the crash. The crash is what the agents do to each other.

What this buys is heterogeneity, thresholds, learning, and the ability to ask what happens if the rules change rather than only if the parameters do. What it costs is calibration, and the book is direct that this is the hard part: the behaviour of the model depends on rules chosen by the modeller, and a sufficiently flexible model can be made to produce almost any result its author expected. Open tools exist — the book names NetLogo and Repast — but the discipline is not in the software. It is in testing whether the emergent behaviour survives changing the assumption you were least sure of.

INNOVATION AS SELECTION

Evolutionary economics treats firms as carriers of routines, the market as a selection environment, and innovation as variation thrown against it. The lineage runs through Nelson and Winter's evolutionary theory of economic change, and the shift it asks for is uncomfortable: firms are not, in this account, optimising. They are repeating what worked with some variation around the edges, and the environment is keeping a subset. Growth at the level of an industry is then not the sum of good decisions. It is differential survival.

The book puts a number on the awkward part. Set the market standard at a quality of 5. An incumbent producing at quality 6 has a fitness of (6 − 5)², which is 1. An entrant producing at quality 8 has (8 − 5)² plus a novelty term, giving 9 + 3k, where k is how strongly the market rewards being new. When k is large enough, the entrant wins. Read the two expressions again: the incumbent's fitness never changed. It was 1 before the entrant arrived and it is 1 afterwards. What moved was the terms of selection.

The incumbent did not get worse. The landscape it was fit for stopped being the landscape.

This is what makes creative destruction, in Schumpeter's phrase as the book uses it, both uncomfortable and structural. The clearing is not waste around the edges of progress; the clearing is the mechanism by which room appears. Set beside it is Arthur's work on increasing returns and lock-in, which establishes the other half of the honest picture: where feedback is strong enough, the surviving standard need not be the better one, only the one that got ahead early enough to be chosen by everyone afterwards. Selection is a filter. It is not a judge.

Which yields the practical reading for anybody building something. The question is not whether the product is good, because good is not a property of the product. It is a relation between the product and a landscape that other people are actively rewriting. A firm can be superb at a fitness function that is quietly being replaced underneath it, and will have no local signal of this until the replacement is complete. Variation is the cheap insurance: a spread of small experiments outperforms one large correct guess, because the landscape gets a vote on what correct means.

ANTIFRAGILE BY DESIGN

Robust and antifragile are different words for different things, and the book keeps them apart. A robust system resists a shock and comes out unchanged. An antifragile one — the term is Taleb's, and the book credits it — comes out better than it went in. The forest fire that clears deadwood leaves ground that grows more than it did before. A firm forced by a supplier's collapse to find and qualify a second supplier now has something it would never have paid for voluntarily, and the next collapse will cost it almost nothing.

Three properties produce the behaviour, and all three are in the book. Diversity: a variety of industries, suppliers, customers, funding sources, so that no single failure mode reaches everything at once. Redundancy: slack, spare capacity, reserves — which look exactly like waste on every day the shock does not arrive, and are the only reason the organisation is still there on the day it does. Loose coupling: connections weak enough that a part can fail locally without dragging its neighbours, where tight coupling is faster, more synchronised, and transmits failure at the same speed it transmits everything else.

The tension with everything a spreadsheet rewards is worth stating flatly. Optimisation removes slack, consolidates suppliers, shortens paths, and tightens coupling. Every one of those steps raises measured return and lowers the system's capacity to absorb a surprise, and every one of them is individually defensible in the meeting where it is proposed. Nobody decides to be fragile. Fragility accumulates, one reasonable improvement at a time, until the structure has no give left anywhere and the first real shock finds that out on everyone's behalf.

Nobody decides to be fragile. Fragility is what a hundred sensible efficiencies add up to.

Antifragility is not automatic, and treating it as a property that systems simply have is the second way this literature gets misread. A system gains from stress only when it can vary, select and retain — when there is something else to try, some way to tell what worked, and a memory that keeps the answer. Strip out any of the three and a shock is just damage with a flattering name. The design question is therefore never how to avoid shocks. It is how to make them informative, and survivable at a scale that does not take the whole structure with them.

WHAT TO DO WITH IT

The book refuses to leave this at the level of admiration, and closes every chapter with a protocol. The first move is to draw the loops for something real and owned — the business, the household, the portfolio — naming which loops amplify, which balance, and where the delays sit. Almost every surprise in a small system turns out to have been one reinforcing loop with a lag that nobody had written down. Drawing it takes an afternoon, and the afternoon is the whole intervention.

The second move is to count exposures rather than assets. The book's argument about diversification is that asset classes are labels while correlations are the thing, and two holdings that answer to the same underlying condition are one holding paying two sets of fees. So the question is not how many positions there are but what condition would hurt each of them, and how many of them share one. That list is usually shorter than the portfolio, which is the finding.

The third is to run experiments that are small, reversible and frequent. Selection needs variation to work on, and a spread of attempts in which most fail cheaply and a few compound is not a timid strategy — it is the only strategy shaped like the landscape it is operating in. Failure stops being a verdict and becomes the price of information, paid in advance and in small denominations.

The fourth is to keep slack and defend it by name. Write down which shock each piece of redundancy exists for, and the note survives the next efficiency review, which is the only review that will ever threaten it.

That is the shape of the argument, and the book runs it properly: ten chapters across 36,864 words, each one carrying a story, the living-systems idea underneath it, the mathematics written out with the numbers substituted rather than gestured at, a worked case in the markets, a protocol for use, and a set of reflection prompts. Every chapter is free to read. Buying the volume is for keeping it — the file on a shelf that is yours, rather than a page that needs a connection.

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

• Emergence and Self-Organization: Foundations of Complex Economic Systemsopen this in the house search
• Networks and Interdependence: Mapping Economic Relationshipsopen this in the house search
• Feedback Loops and Nonlinear Dynamics: Understanding Cycles and Crisesopen this in the house search
• Agent-Based Modeling: Simulating Heterogeneous Economic Actorsopen this in the house search
• Innovation, Entrepreneurship, and Evolutionary Economicsopen this in the house search
• Financial Markets as Complex Adaptive Systems: Volatility, Bubbles, and Crashesopen this in the house search

A forecast assumes the future is being delivered. A living system is busy negotiating it.

Emergence is what a system does that no member intended and no member could have done alone.

Every crisis is a measurement. Fragile systems pay for it and learn nothing.

Optimise a structure far enough and you have bought return by selling the ability to be wrong.

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.