Haute Lumière · The reading

The economy is not a machine, and the models that treat it like one keep being surprised

A working method for modelling something that is alive: draw the boundary, count the stocks, write the loops, and let the data refuse you.

She has been on the same page for twenty minutes. Nothing is wrong yet; she is deciding where the edge of the thing goes.

THE WRONG PICTURE

The picture underneath most economic forecasting is a machine. Inputs go in one side, outputs come out the other, and the arithmetic in between is assumed to hold still while somebody does it. That is a serviceable picture for a conveyor belt and a poor one for anything that adapts while it is being measured. A machine does not change its behaviour because a forecast about it was published. A market does that before lunch.

The alternative picture is a forest. Water nourishes the trees, the trees hold the soil, the soil holds the water, and no part of that arrangement can be understood by taking it out and weighing it. A tree is not an object so much as a set of rates — sunlight in, sugar made, water drawn up, oxygen released, leaves dropped and returned to the ground as nutrients. Stop the rates and there is no tree left to study. An economy has the same property and hides it better, because balance sheets are printed as though they were photographs when they are closer to weather reports.

The question is never whether the arithmetic is correct. It is whether the picture underneath the arithmetic is of something that can move.

Six words carry the rest of the method, and they are worth having in hand before anything is drawn. Flows are the constant movements — money between accounts, goods across borders, information through networks. Stocks are the accumulations where flows gather and sit: cash reserves, infrastructure, biomass, skills, debt. Feedback is what happens when a flow changes a stock and the stock then changes the flow, which is the mechanism behind almost every behaviour anyone finds surprising.

The last three are the ones that separate a living picture from a mechanical one. Coupling asks how tightly the parts are bound: a drought in one region reaching food prices globally is strong coupling, and an economy with one dominant industry is more tightly coupled, and more fragile, than a diversified one. Emergence is the observation that a flock of birds turns in formation without any bird holding the plan, and that market trends arise the same way, out of millions of independent decisions none of which contain the trend. Antifragility is the property that a forest fire clears deadwood and a recession forces adaptation — that some systems come back stronger from the thing that was supposed to end them.

None of this is a rejection of ordinary economics. It is an insistence that the model carry the same shape as the thing it is modelling, because when the shape is wrong the errors are not random. They are systematic, they all point the same way, and they arrive at the worst possible moment.

WHERE IT ENDS

Every system has a boundary, and the first real decision a practitioner makes is where to put it. For a forest it might be the edge where the trees give way to grassland. For a company it might be the legal entity. For a household it is the people who share the money. None of these edges is sharp; all of them are permeable, with influence crossing in both directions. That is not a defect in the definition. It is the reason the definition has to be stated out loud.

A boundary that is never declared gets drawn anyway, silently, by whatever data happened to be available. That is how an analysis of a regional economy ends up being an analysis of the three employers who publish quarterly figures. The boundary was set by the filing calendar of three companies, nobody chose it, and nobody can argue with it afterwards because nobody knows it is there.

Consider the beekeeper at the market with his jars of honey. Ask him about the honey and he will talk about the bees — the hive, the dance they use to tell each other where the nectar is, the fine hairs on a worker bee's legs that trap pollen. Ask what makes the jam-maker's blueberry lavender good and he will name the sunshine that ripened the berries, the bees that pollinated the blossom, the soil, and four generations of people who learned when to stop stirring. Neither man is being poetic. Both are naming the boundary of the system that produces the thing on the table, and both boundaries are wider than the product.

A boundary drawn too tight produces a model that is right about a small thing and wrong about the reason it happened.

Once the edge is set, three questions follow and they are always the same three. What accumulates inside it — the stocks. What moves across it and between the parts — the flows. Who acts inside it — the actors, which in a forest means trees, fungi, animals, bacteria and the people with chainsaws, and in a regional economy means firms, agencies, households and whatever regulator can change the rules on a Tuesday.

Running those three questions on a technology company gives something immediately usable. The stocks: market capitalisation, which is the accumulated judgement of investors, and research investment, which is the accumulated capacity to still be interesting in four years. The flows: revenue in, expenses out, investment inflows and outflows as people buy and sell the shares. The actors: competitors whose product launches change the demand curve, customers whose preferences move with technology and fashion, and regulators. That list is not yet a model. It is the thing without which a model is only an opinion with decimal places.

STOCKS AND FLOWS

A bathtub explains more of this than a textbook does. The water in the tub is a stock. The tap is an inflow, the drain is an outflow, and the level rises whenever the tap beats the drain. That single sentence, written properly, is the foundational equation of the entire method: the rate of change of a stock equals inflow minus outflow. Everything else is that equation applied to something that matters more than bathwater.

The distinction sounds trivial until it is used to sort real quantities, at which point it starts catching errors. A savings account is a stock; a salary is a flow. A nation's accumulated wealth is a stock; its trade is a flow. A company's inventory is a stock, subject to an inflow of production and an outflow of sales. Confusing the two is the most common analytic mistake there is, and it has a signature: somebody celebrating a rate of change as though it were a level, or panicking about a level as though it were a rate.

Take a lake. The boundary is the shoreline and the vegetation at the edge. The stocks are the volume of water, the fish population, the algae biomass. The flows are rainfall adding water and evaporation taking it, fish reproducing and birds eating them, algae growing through photosynthesis and being consumed. Written for the fish, one line holds all of it: the population next period equals the population now, plus births, minus deaths. That is not a simplification of the ecology. It is the ecology, stated in the only terms that let anyone reason about a stocking programme or a fishing licence without arguing from instinct.

Almost every disagreement about an economy is two people describing the same stock at two different moments and calling it a trend.

The same line, moved into finance, changes only its nouns. Market capitalisation next period equals market capitalisation now plus net investment inflows. Research capacity next period equals research capacity now, plus new spending, minus the depreciation of what was funded three years ago and has since been overtaken. A firm spending two hundred million a year and losing a tenth of its existing research stock annually is running up an escalator, and the equation makes the escalator visible in a way an annual report does not.

For a household the nouns change again and the structure does not. The stocks are savings, assets and debt. The flows are income, spending, interest and repayment. Once the household is drawn this way, questions that felt like character judgements become questions about rates: which flow is largest, which one is easiest to change, which stock does the change actually reach. That is the whole promise of this way of working. It moves the argument off the person and onto the plumbing.

The column on the left is history and the column on the right is a guess. Most of the work is in keeping them apart.

THE LOOPS

Feedback is the point where the method stops describing and starts explaining. A flow changes a stock; the stock, having changed, changes the flow. That circle is the reason systems behave in ways that no list of their parts predicts, and it comes in exactly two flavours regardless of what is being modelled.

Reinforcing loops amplify. Consumer confidence rises, spending rises, production and employment rise, confidence rises further. A company invests in research, its capability grows, new products raise revenue, revenue funds more research. These are the loops behind anything that seems to happen suddenly, because a reinforcing loop is slow while the quantity is small and violent once it is large, and nothing about the mechanism changes in between. What changed was the size of the thing being multiplied.

Balancing loops resist. Wages rise, the cost of living rises with them, demand cools and the pressure comes off. Predators multiply, prey become scarce, predators decline, prey recover. A thermostat cools the room it is in. These are the loops that keep systems inside a range, and they are also the loops that quietly defeat interventions — push on a system with a strong balancing loop and the system pushes back with exactly the force required to end up where it started, which is experienced by whoever pushed as a failure of will.

A reinforcing loop is not a trend. It is an arrangement, and arrangements can be changed by anyone who can see them.

Most interesting behaviour comes from the two acting together on different timescales. A firm chasing quarterly profit cuts training. Productivity falls and the best people leave, which is a balancing loop nobody wanted, arriving eighteen months after the decision that caused it and long after the decision has stopped being discussed. Research spending raises capability and also drains cash reserves; whichever of those two loops is faster determines whether the policy reads as visionary or reckless, and the same policy can be both depending on the balance sheet it is run from.

Coupling decides how far a disturbance travels. In a tightly coupled system a shock anywhere appears everywhere, which is efficient in good weather and catastrophic in bad. Loose coupling absorbs — a shock stays local, the rest carries on, and the system as a whole gets slower and harder to kill. This is why diversification in a portfolio and diversification in a national economy are the same move, made at different scales, for exactly the same reason.

Drawing the loops is not optional decoration on top of a stock-and-flow diagram. It is the step where the diagram acquires the ability to be wrong about something specific, which is the only kind of wrong that is any use.

THE LIMIT TERM

There is one piece of mathematics worth learning properly, because a surprising share of real behaviour is a special case of it. Start with unlimited growth: the rate of change of a quantity is proportional to how much of it there already is. Written out, that is dX/dt = rX, where X is the quantity, t is time and r is the intrinsic growth rate. A colony of bacteria doubling every hour is this equation with r set to one, and it produces the familiar shape — a hundred, two hundred, four hundred, and shortly afterwards a number that has stopped meaning anything.

Nothing does that for long, because the world is finite and the resources run into a ceiling. The logistic equation adds exactly one term to say so: dX/dt = rX(1 − X/K), where K is the carrying capacity, the largest quantity the environment can sustain. The new bracket is a brake. When X is small relative to K the bracket is close to one and growth is effectively exponential. As X approaches K the bracket shrinks towards zero and growth stops on its own, without anybody deciding it should.

Numbers make the brake visible. Take a deer population with a carrying capacity of 500 and an intrinsic growth rate of 0.2 per year. At 100 animals the model gives 16 deer per year. At 116 it gives about 17.8 — still climbing, but the increase in the increase is already flattening. Take a rabbit population with a carrying capacity of 100 and a growth rate of 0.5 per year: at 10 rabbits the rate is 4.5 per year. In every case the population is expanding and the expansion is being metered by how close it already is to the limit.

The interesting parameter was never the growth rate. It is the ceiling, and the ceiling is the thing nobody puts in the deck.

The reason to know this equation is that it moves. Replace population with output and K becomes the maximum sustainable production given resource constraints. Replace it with price and K becomes market equilibrium. Replace it with biodiversity and K is what the habitat will carry. The same two terms — one that wants to grow, one that says how much room is left — describe a species, a product category and a balance sheet, which is a strong hint that the shape is real rather than borrowed.

Growth rates are easy to estimate and get quoted confidently. Carrying capacity is hard, contested and usually guessed, and it is also the term that decides everything about the second half of the curve. Any projection that names r to two decimal places and leaves K implicit is not a forecast of the system. It is a forecast of the system's youth, extended past the point where youth was the relevant fact. Solving these equations by hand is rarely possible; the practical route is numerical, with something like Python and SciPy stepping the system forward, which is a detail of execution rather than of understanding.

DIRTY DATA

Then there is the spreadsheet. Rows upon rows, some neatly formatted and some apparently entered by a caffeinated squirrel. Half the cells are blank. Others carry codes whose meaning left with the person who invented them. And scattered through it are outliers so far out of range they look like they arrived from another dimension. This is the normal condition of data, and any method that assumes otherwise is a method that has never been run on anything.

The honest way to think about it is archaeology. The practitioner is excavating the past of a company, a region or a community, and the finds are scattered across databases, reports, regulatory filings and occasionally handwritten notes. Some are clean — a quarterly revenue table, a census return. Most are oblique: a line in meeting minutes hinting at a strategy that was never announced, a small news item revealing a relationship between two parties who claimed not to have one. Each is a fragment, and the work is assembling them into something that holds.

Cleaning has three parts and none of them are clerical. Units must be made consistent, which sounds beneath mention until half the factory records are in feet and half in metres and the trend line goes vertical in March. Missing values must be handled, either by estimating them from what is present or by excluding them, and whichever is chosen has to be written down where a reader can find it. Outliers must be investigated rather than removed — one month of unusually high production might be a holiday rush, in which case it is the most informative row in the file, or it might be a typo, in which case it is nothing. Only knowledge of the system tells them apart.

Deleting an outlier before understanding it is the fastest way to build a model of the months when nothing happened.

Then the gaps that will not close. Perhaps the supply relationships of a particular manufacturer are not disclosed anywhere. Perhaps the regulatory direction for the next three years is genuinely unknown, because it depends on an election. The temptation is to fill these with a confident central estimate, and the discipline is to refuse. A gap named is a gap a reader can reason about. A gap filled quietly becomes indistinguishable from a measurement, and it is exactly the number that gets quoted back in a meeting a year later.

The technique for handling the ones that will not close is scenario planning: build the analysis under several sets of assumptions and look at the range of outcomes rather than the middle of them. This is not a fudge. It is a more accurate description of what is known, and it is more useful to whoever has to decide, because the spread of the outcomes is the actual risk and the point estimate never was.

Data in this method is not a snapshot. It is the visible residue of flows that are still running — every transaction, every interaction, every input and output leaving a trace. Read that way, even an incomplete file tells you where the energy is moving and where something has jammed.

Nothing here is finished. A model is kept the way a garden is kept, and the keeping is the method.

DIAGRAM TO EQUATION

At some point the whiteboard covered in sticky notes and arrows has to become something a computer can step forward, and this is where most systems thinking quietly stops. It should not. The translation is mechanical once the diagram is honest, and doing it is what converts an interesting way of talking into a method that can be checked.

Start with the smallest real example: a population limited by a resource. Let P be the population and R the available resource. Population follows the logistic form already given, growing at rate r and braking as it approaches what the resource will carry. Resources follow their own line — they replenish at some rate b and are consumed in proportion to the population, so dR/dt = b − cP, where c is consumption per head. Two equations, four parameters, and a system that can now be simulated rather than argued about.

Add a third stock and the structure holds. Suppose the population generates waste at a constant rate per head, so dW/dt = gP. Now waste accumulates, and if waste reduces the resource or harms the population, two new loops appear — neither of which was visible in the prose description, both of which were implied by it. This is the practical value of the translation. Writing the equations does not make the model true. It makes the assumptions explicit, and an explicit assumption is one that a colleague can disagree with on Tuesday morning.

A diagram can be admired. Only an equation can be contradicted, which is why the equation is the one worth having.

Relationships that are not stocks and flows get written as plain functional statements. Demand against price can be as simple as a line: a quantity of 100 falling by 2 for every unit the price rises. Production against inputs can use the Cobb-Douglas form, Y = A·K^α·L^(1−α), where Y is output, A is total factor productivity, K is capital, L is labour, and α says how much output responds to capital rather than to people. A toy manufacturer with monthly production figures, factory floor area and headcount can estimate A and α by regression and learn something it could not learn by reading its own accounts: whether it is a company that grows by buying machines or a company that grows by hiring.

Those estimates arrive with uncertainty attached, and the statistical software will hand over confidence intervals alongside them. Keep the intervals. They are not a caveat on the result, they are part of the result, and a parameter reported without its interval has been rounded into a claim it cannot support.

The aim is not a mirror of reality. Living systems are too dynamic for that and always will be. The aim is to capture the essential relationships well enough that the model responds to a change the way the real system does, and then to keep adjusting it as that stops being true.

EARNING TRUST

A model that has never been tested against the world is a ship in a bottle: intricate, admirable, and structurally incapable of going to sea. The story the book tells about this is of a self-described pioneer in predictive gastronomics whose equations predicted a fifteen per cent surge in avocado toast consumption among millennials — mathematically sound, internally consistent, and describing a population that would have been eating avocado toast for breakfast, lunch and dinner. The mathematics was not the problem. Nothing had ever told the mathematics it was wrong.

Calibration is the first correction. Gather the history the system actually produced, run the model against it, and adjust the parameters until the model's output tracks what happened. The image is a telescope: turning the focus until a distant object comes into view. Running this on an organic farming cooperative means feeding in past land acquisitions, labour hours, yields and prices, and discovering, for instance, that the assumed efficiency of labour has to be revised because the historical trend says so.

Calibration alone proves very little, because a model with enough free parameters can be bent to match any history at all. Validation is the part that matters: test the calibrated model against data it has never seen. Hold back a quarter, predict it, and compare. A model that reproduces the past perfectly and then misses the held-back quarter badly has not learned the system, it has memorised the file.

Any model can be made to agree with the past. Only a model that survives data it was not shown has told you anything.

Errors are expected and are not the verdict. The cooperative's model predicted a fifteen per cent increase in yield where the actual increase was twenty. That gap is not a failure; it is the most informative output of the exercise, provided the next three questions get asked. Is the error consistently in one direction, which would mean something structural is missing. Does it cluster around particular variables. Does it appear only under particular conditions, such as unusual weather or an unusual price. Answer those and the model improves. Skip them and the discrepancy gets rounded off as noise, which is how a systematic blind spot becomes permanent.

The magnitude and the character of the error are two different findings, and the character is worth more. A model that is off by twenty per cent at random is usable with wide intervals. A model that is off by five per cent always in the same direction is broken in a way that will not stay at five per cent, because whatever is missing from it is still there, still operating, and will be larger next year.

Calibration and validation are not a stage that ends. They are a habit, repeated as new data arrives, and the model that results is not a static representation but something kept in working order alongside the system it describes.

IN THE MARKETS

Applied to an actual position, the method changes what gets counted rather than how the arithmetic is done. Consider an analyst looking at a solar technology company promising higher panel efficiency than anything on the market. The mechanical approach projects a growth rate and discounts the cash flows. The living-systems approach first names the three loops the company's future actually sits inside: technological innovation, whether the efficiency claim survives testing; market adoption, which depends on price, competition and public awareness; and manufacturing scale, which depends on raw materials, supply chain and the capacity to build capacity.

Then the loops get quantified, starting with one. A company entering at a negligible 0.1 per cent market share and growing that share at 25 per cent a year reaches 0.125 per cent after one year, 0.244 per cent after three, and 0.488 per cent after five. Written out like that, the projection does something a slide never does — it shows that five years of impressive compound growth still leaves the company with under half a per cent of its market. Whether that is a triumph or a disappointment depends entirely on what was promised, and now both parties can see the same number.

Sovereign debt yields to the same treatment. A simplified relation between bond price, yield, country risk premium and the global interest rate — price as yield plus risk premium, over the benchmark rate — makes the dependencies explicit rather than intuitive. With a global rate of 3 per cent, a yield of 6 and a risk premium of 2, the relation gives 2.67. More usefully, it says which way the price moves when each input moves, and where the loops run: global rates rising makes the bond less attractive and the price falls; domestic policy that reduces the risk premium raises it. The number is a by-product. The structure is the analysis.

A point forecast is a scenario that has forgotten it was one.

Scenarios are where the method pays. Take a solar market growing at 15 per cent a year on a base case, with panel costs falling 8 per cent annually, supportive policy and rising consumer awareness. An optimistic case where technological breakthroughs accelerate cost reduction takes adoption to 20 per cent annually. A pessimistic case where political instability reverses policy support slows it to 5 per cent. Three internal rates of return, three present values, and a decision made against a range instead of against a single number that was always going to be wrong by some amount nobody had estimated.

The same lens explains why some companies outperform their line items. A clothing company known for durable goods and environmental commitment holds stocks that are not on the balance sheet at all: brand reputation, product quality, and a genuinely small environmental footprint. Reputation attracts customers, customers produce revenue, revenue funds research into better materials, better materials strengthen the product and the reputation. That is a reinforcing loop, it is the actual asset, and a valuation that counts only inventory and receivables will keep concluding the shares are expensive while the loop keeps running.

AT THE DESK

The protocol is six steps and it is the same six whether the system is a sovereign fund or a household. Define the system and its boundary. Identify and quantify the key stocks, in real units, from real data. Define the flows that change them. Write the relationships between flows and stocks as equations. Simulate, vary the parameters, and look at how the system responds. Then iterate, because the first version is a draft and treating it as anything else is the error the whole method exists to prevent.

Run on a household it is unglamorous and immediately useful. The system is the people who share the money. The stocks are savings, assets and debt; the flows are income, spending, interest and repayment. The data is bank statements and card bills, which need the same cleaning as any other source. What emerges is not a budget but a set of loops: spending above income raises debt, debt raises interest, interest raises required spending, and the loop turns whether or not anybody has noticed it turning. Seeing it stated as a loop is different from being told to spend less, because a loop has several places where it can be interrupted and only one of them involves willpower.

Run on an institution, the same six steps identify where an intervention would actually land. Most proposals arrive aimed at a stock — raise this, cut that — when the stock only moves at the speed of the flows attached to it. The point where a system actually gives is nearly always a flow or the strength of a loop, and the diagram is what makes the difference visible before the money is committed.

Aim at a stock and you get a number for one quarter. Aim at a loop and you get a different system.

Which leaves the shocks. The instinct is to model for stability and treat every disruption as damage, and the living-systems view is less sentimental than that. A forest fire clears deadwood and lets new growth through. A downturn forces firms to adapt in ways that no comfortable year ever produced. Some quantity of stress is a condition of health rather than a threat to it, and a system engineered for perfect smoothness is usually a system that has quietly removed everything that would have helped it bend.

So the design question changes. Not how do we prevent every disturbance, which is unachievable and expensive in the attempt, but where does this system store slack, how loosely are its parts coupled, how many independent sources does it draw on, and what does it gain from being shaken. Diversified holdings and a social safety net are the same answer at two scales, and both look like inefficiency right up until the week they are the only reason anything is still standing.

That is the argument, end to end: draw the boundary and say it out loud, count what accumulates and what moves, find the loops, write them down as equations, get the data and be honest about its holes, calibrate against history, validate against what the model has never seen, and keep the thing in working order as the system changes underneath it. The whole book is on the shelf and every chapter of it is free to read — thirteen chapters, worked through with the numbers left in, in the order a practitioner would actually meet them.

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

• The Method, End to Endopen this in the house search
• Framing the System: Boundaries, Stocks, Flows, and Actorsopen this in the house search
• Getting the Data: Sources, Cleaning, and Honest Gapsopen this in the house search
• Building the Model: From Diagram to Equationsopen this in the house search
• Calibration and Validation: Making a Model Earn Trustopen this in the house search

A boundary is not discovered. It is declared, and then paid for by everything the model gets wrong.

Every forecast is a confession of what its author believes cannot change.

A model that has never been refused by data has never been tested, only admired.

Uncertainty priced honestly is worth more than precision invented.

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.