When markets fail

An externality is a cost or a benefit that lands on somebody who was not party to the trade. The seller counts their own costs when they set a price, and the buyer counts their own gain when they pay it; neither has any reason to count the neighbour. So the neighbour's cost is real, and it is nowhere in the price. The market then does exactly what the price tells it, which is to make more units than anybody would choose if the whole bill had been on the tag. That is the hatched band in the figure, and it is the honest definition of "the market failed": not that the mechanism broke, but that the message it was acting on was incomplete.

The same thing runs the other way. A positive externality is a benefit that lands outside the trade: a vaccine protects the people the patient sits next to, a well-documented internal dataset saves every team that later reuses it. Here the market makes too little, because the person paying for the unit gets only part of its value. Whoever bears the cost or the benefit that the price missed is the person with a claim, and the whole of environmental policy, and most of data-platform policy, is an argument about how to put that missing line back on the tag: a tax, a cap, a chargeback, a rule.

The data habit: whenever a decision looks cheap, ask whose cost is not in the number. A query that runs in seconds because it saturates a shared cluster is cheap to the analyst and dear to everyone queued behind it. A model shipped without a monitoring budget is cheap to the team that shipped it and dear to the team that inherits it. Both are the hatched band.

A public good is the extreme case of a positive externality: nobody can be excluded from using it, and one person's use does not use it up. A lighthouse, a clean data dictionary, a public benchmark. Because nobody can be excluded, nobody will pay first, since the plan of waiting for somebody else to pay is always cheaper. So the market makes very little of it, and the things that get made are made by whoever is willing to pay for everybody, which is usually a government, a foundation or a platform that benefits indirectly.

A commons is the mirror image: a shared resource that anybody can use and that does get used up. A fishery, a shared compute pool, an on-call rota. Each user gets the whole benefit of one more unit and pays only their share of the depletion, so each one uses too much, and the resource collapses even though every user could see it coming. The tragedy is not stupidity; it is that the private cost and the social cost are different lines, and each person is reading the lower one.

Regulation, in the narrow economic sense, is the set of tools for moving the price back onto the higher line: a tax equal to the missing cost, a cap with tradable permits, a quota, a liability rule that makes the seller pay the neighbour, a chargeback that bills the team for the cluster time it actually used. Each tool has its own failures, but the reason any of them exist is the same hatched band.

Start with what the buyer can see: the make, the year, the mileage, the paint. Now what they cannot: whether the engine was ever run dry, whether the service history is real. The seller knows both halves. That gap is information asymmetry, and its first consequence is that the buyer prices the car as a draw from the pool rather than as the specific car in front of them. The offered price is the average of what the pool is expected to contain.

The second consequence is the one that bites. An average price is a good deal for the owner of a lemon and a bad deal for the owner of a sound car, so the sound cars are withdrawn and the lemons are offered. The pool the buyer is drawing from is now worse than the one they priced, so the next offer is lower, which withdraws the next tier of sound cars. Adverse selection is the name for a pool that gets worse because of the price you set for it. In the limit the market for sound cars stops existing, not because anybody lied, but because the honest owners could not prove they were honest.

Akerlof's paper, "The Market for Lemons", Quarterly Journal of Economics 84(3), 1970, pages 488 to 500, works this through for cars, insurance and credit in developing markets, and its closing sections are about the institutions that stop the spiral: brands, chains, licences and guarantees. Every one of them is a way for a good seller to make quality visible before the money moves.

Adverse selection happens before the trade: the hidden fact is a quality that already exists, and the price sorts who shows up. Moral hazard happens after: the hidden fact is a choice the other party makes once the deal is signed, and the deal itself changes the choice. A driver with full insurance parks less carefully. A vendor paid a flat fee for a dataset has less reason to keep it clean than one paid per row that passes a check. A team whose model is judged only at launch has less reason to keep monitoring it than one judged at month six.

The two need different fixes. Adverse selection is fixed by making quality visible up front: a sample, an audit, a signal. Moral hazard is fixed by keeping some of the payoff tied to behaviour after signing: a deductible, a retention clause, a service level, a payment that follows the metric rather than preceding it. Confusing the two is common, and the usual symptom is a beautifully audited purchase that decays in quality within a quarter because nothing after the signature was tied to keeping it good.

A signal is anything a good seller can do cheaply that a bad seller can only do dearly. A warranty is the classic case: a seller who knows the car is sound can promise to fix it for a year at little expected cost, while a seller of a lemon would face a real bill, so only the sound cars get warranties, and the buyer can read the warranty as a fact about the engine. A brand works the same way, since a brand is a promise that is expensive to break. So does a certificate, a published audit, a money-back sample period, or a benchmark score that the seller did not choose the benchmark for.

The test for a signal is cost asymmetry: if the bad seller can send the same signal at the same cost, it is not a signal, it is decoration. A "quality assured" badge that anybody can print is decoration. An independent audit with a named auditor is a signal. A benchmark that the vendor picked, ran and reported is somewhere between, and reading which is the job of the deep-dive track's eighth session, which applies all of this to models and the evals that vouch for them.

Screening is the same trick from the buyer's side: the buyer designs a choice that only the good sellers would accept. A contract that pays per row that passes a check screens out vendors who know their rows will not pass. Insurers, lenders and procurement teams do this all day, and the shape of the contract is itself a statement about what the buyer suspects.

A measure is a number chosen because it moves when something you care about moves. Ticket resolution time moves when support gets better. Test-set accuracy moves when a model gets better. The number is useful precisely because of that correlation, and the correlation exists because nobody is trying to move the number for its own sake. A target is the same number with a reward attached. The reward gives everyone who can touch the number a reason to move it, and the cheapest way to move a number is almost never the way that moves the thing behind it.

Resolution time falls when agents close tickets early and reopen them. Test-set accuracy rises when the test set leaks into training, or when the team tunes against it for a year until it is a validation set in all but name. Neither of these is fraud; each is a rational response to the price the target put on the number. And in every case the number keeps rising while the thing it stood for stops moving or falls, which is the hatched wedge in the figure and the most expensive shape in management.

Three defences, none of them complete. Pair the metric with a counter-metric that the same cheap move would push the wrong way: resolution time with reopen rate, accuracy with a held-out set nobody has seen. Rotate the measure before it is fully learned. Audit the mechanism, not the number: once a quarter, trace a sample of the units behind the metric back to the thing, and see whether the number still means what it meant when it was chosen.

The sentence people quote, "when a measure becomes a target, it ceases to be a good measure", is from Marilyn Strathern's 1997 article in European Review 5(3), pages 305 to 321, on audit in British universities. It is usually attributed to Charles Goodhart, whose 1975 paper on UK monetary policy made the same observation about a statistical relationship collapsing once it was used as a control target, in language aimed at central bankers rather than managers. Both are worth reading; the crisp sentence is hers, the priority is his, and the mechanism is the same in both.

The reason the attribution matters beyond tidiness is that the two texts point at different fixes. Goodhart's context says: a relationship that held under one policy will not hold under a policy that leans on it, so do not build the policy on the relationship. Strathern's context says: an audit culture changes the audited, so measure with care about what the measuring does. A data team needs both readings, one for the model that predicts a metric and one for the dashboard that rewards it.