You look at a screen. Gold is one number, oil is another, bitcoin is a third. They sit in a neat column with small arrows beside them, and they look like facts of the same kind as the tem­per­a­ture. Says who?

A barrel of oil weighs about 136 kilo­grams. Gold has an atomic number of 79. At 07:24:04 on the morning of 16 September, a bitcoin on Binance cost $75,807.20. English grammar makes those three sen­tences sound alike. The barrel would weigh the same on an island with nobody on it. Gold had 79 protons before anybody counted them. The price of a bitcoin was true only because, at that instant, some­body was pre­pared to sell one at that figure and nobody had yet taken them up on it.

A price needs a sur­pris­ing amount of company. It needs a buyer and a seller. It needs a place where they can find each other, and rules for what happens when they do. It needs a unit to be written in, a clock to be

what­ever quan­tity happens to be resting there, and not one unit more.

In our second the median print was 0.0077 bitcoin, about $583. At that size the price on the screen and the price paid are the same number. The largest print was 6.51994 bitcoin on Binance, 67 mil­lisec­onds in, worth $494,258. In the last snap­shot of Binance's book before the second, 6.83704 bitcoin were offered at exactly $75,807.20, so that trade fitted inside the best offer and paid the screen price. A buyer a few bitcoin larger would have paid more for the last of them.

741 prints crossed in that second. The smaller half, 370 trades, carried 1.46% of the money. The largest seven, one per cent of the count, carried 29.9%. The tape is mostly trans­ac­tions that move almost no value, punc­tu­ated by a few that move nearly all of it, and every one of them prints in the same size of type.

The small­est trade we cap­tured was 0.00000003 bitcoin on Coinbase, 896 mil­lisec­onds in: about a quarter of a cent, real, settled. On a chart it counts exactly as much as the half-million-dollar one, because a chart of prices has no column for size. Act II is about that missing column.

What the screen leaves out

So one quote carries a lot it never says. It has a unit, which may not be the one you assume. It has an age, which is never dis­played. It has a size, which is almost never dis­played. It comes from one venue, or from a rule that com­bines several. And it sits on a stack of deci­sions made by people and pro­grams who are not visible at all.

Nobody is hiding any­thing. A screen with every foot­note would be unread­able, and whether bitcoin was $75,807 or $75,844 at break­fast changes nothing for most people. The trouble lives at the edges, and so does the rest of this issue: large sizes, thin books, con­tracts, and the rare hours when a market stops and a number that was a price a minute ago stays on the screen as dec­o­ra­tion.

The dis­ci­pline this issue asks for is small. Whenever a price matters, ask five ques­tions of it. In what unit? From where? How old? Good for how much? Made by what rule? Most of the time the answers are boring. When they are not, the number was never telling you what you thought.

So what was the price?

By the end of the second all five venues had moved up, none by the same amount. The Binance offer that opened at $75,807.20 closed at $75,844.01. Ask at the start of the second what a bitcoin cost and there were five true answers. Ask at the end and there were five dif­fer­ent true answers. Nothing about bitcoin had changed. Plenty about the market had.

Weight belongs to the object. A price is a result: of who was willing to act, where, in what unit, at what size and when, and of who reported it, after how long. The rest of Act I stays with the number itself. It asks what happens when a price is cal­cu­lated instead of traded, when it goes below zero, and when some­body has to decide, offi­cially, which number counts.

Price is not some­thing an object has. Price is some­thing a market does.

There are prices at which nobody bought any­thing. They are not rare, and they are not errors. Some of the most impor­tant numbers in finance are of this kind, and most people who rely on them have never noticed.

An index level is one. So is the mark price an exchange uses to decide whether your futures posi­tion has run out of margin. So is the net asset value of a fund, struck once a day from some­body else's prices. So is a volume-weighted average price, the bench­mark a broker is judged against, and a fair value an auditor signs off for an asset that has not traded in months. Each is com­puted. None needs a buyer and a seller to have met at that figure.

Here is one from our second. Take every print on the four tether venues between 07:24:04.000 and .999, weight each by its size, and average. The answer is $75,822.48. It is a per­fectly rea­son­able

reading them, because they are the dif­fer­ence between two numbers that both claim to be the price of the same thing.

CF Benchmarks, which cal­cu­lates the bitcoin ref­er­ence rate CME futures settle against, is a useful example of how spe­cific those rules get. Its daily rate takes trades from a set of con­stituent exchanges over one hour, splits the hour into twelve five-minute par­ti­tions, takes a volume-weighted median in each, and aver­ages the twelve equally. No trade needs to have hap­pened at the result. Several billion dollars of futures settle on it anyway.

Our own number

We should declare our inter­est here, because our own systems make con­structed prices, and our capture of this second con­tains one.

One of our col­lec­tors records a ref­er­ence price for bitcoin every ten seconds, for risk checks. For 07:24:04 it recorded $75,807.195. The source field reads binance-fallback: the oracle we would prefer was not the one used. The figure is the mid­point of Binance's best bid and ask at the start of the second, and it has half a cent in it. Binance quotes in whole cents, so no trade could ever have hap­pened at that price. The row reached our data­base 1.227 seconds after the instant it described, by which time Binance's best offer was $75,844.01.

Nothing about that is broken. The row is labelled hon­estly: where the number came from, when it was pub­lished, how old it was on arrival, and whether it passed our checks. Often the most useful price has no trans­ac­tion behind it at all, only a state­ment with its prove­nance attached, and it deserves exactly as much trust as that prove­nance is com­plete.

The accoun­tants got there first

Accounting stan­dards settled this argu­ment long ago, and their answer is worth bor­row­ing. Both the inter­na­tional and the American rules for fair value sort every mea­sure­ment into three levels. Level 1 is a quoted price in an active market for an iden­ti­cal asset: a traded price, used as it stands. Level 2 uses observ­able inputs that are not quite that, such as quotes for similar assets or prices from markets that trade rarely. Level 3 uses inputs that cannot be observed at all, and relies

On 20 April 2020, oil did not merely become worth­less. For some con­tracts, getting rid of it became valu­able.

It is 2:30 in the after­noon in New York. The May con­tract for West Texas Intermediate crude on the New York Mercantile Exchange settles at minus $37.63 a barrel. On Friday it settled at $18.27. Today it has traded as low as minus $40.32. Tomorrow, its last day of trading, it will expire at $10.01. Right now, anyone holding a con­tract to receive a thou­sand barrels of oil in May can get out of it only by paying some­body else nearly $38,000 to take it.

The head­lines said the price of oil had gone neg­a­tive. That is almost right, and the part that is wrong is the whole story.

What the con­tract actu­ally was

Forget oil in general. A WTI futures con­tract is an agree­ment to take or make deliv­ery of a thou­sand barrels of a spe­cific grade,

during a spe­cific month, at a spe­cific place: Cushing, Oklahoma, a town of tanks and pipe­lines where much of the crude in the middle of the United States passes through. Most people who trade the con­tract never intend to see the oil. They sell or roll their posi­tion before it expires. The ones still holding at expiry either have some­where to put a thou­sand barrels, or find some­body who does.

In April 2020 the world had stopped driving and flying. Demand col­lapsed faster than pro­duc­tion could be shut in, and the oil had to go some­where. Storage every­where was filling, and the market's fear was that the tanks at Cushing would be full by May. The US Commodity Futures Trading Commission's staff later described an over­sup­plied market, an unprece­dented fall in demand, and concern about the avail­abil­ity of storage, all arriv­ing at once on the day before expiry.

For a holder of the May con­tract with nowhere to store the oil, the cal­cu­la­tion on the after­noon of 20 April had stopped being what oil was worth. It had become what it would cost to avoid receiv­ing it. Past a certain point, paying a stranger to take the obli­ga­tion was the cheap­est option avail­able. The CFTC staff found the fastest part of the fall hap­pened between 1:00 pm and the 2:30 pm set­tle­ment.

It was never the price of oil

On the same after­noon, the June con­tract for the same grade at the same place traded above $20. Oil a month later was worth a pos­i­tive amount. Oil that had to be taken now, with no tank to put it in, was worth less than nothing. Same crude, same town. What dif­fered was time, place and a tank: the three things the word price usually hides.

The minus sign was a truth­ful report of a phys­i­cal problem. The barrels had not lost their energy content. What had gone neg­a­tive was the right to receive them at a moment when receiv­ing them was a lia­bil­ity. A futures price describes a con­tract, and a con­tract describes an obli­ga­tion, and an obli­ga­tion can be worth avoid­ing.

Other ways below zero

Electricity goes below zero rou­tinely now. Germany recorded 457 hours of neg­a­tive whole­sale prices in 2024, up from 301 in 2023, accord­ing to its Federal Network Agency, mostly around midday in summer when solar output peaks. Power cannot be stored cheaply at scale. Some plants are expen­sive to switch off and on again, and some gen­er­a­tors are paid for output under support schemes regard­less of the market price. At those hours, paying some­body to consume a megawatt-hour is cheaper than the alter­na­tives, and the price says so.

Negative inter­est rates look similar and work dif­fer­ently. The European Central Bank set its deposit rate below zero in June 2014 and kept it there until July 2022. The Bank of Japan's policy rate was neg­a­tive from 2016 until March 2024. Those were admin­is­tered rates, chosen by central banks as policy, and what they priced was the cost of holding money at the central bank. They are rel­a­tives of the oil episode, cousins

If prices can be con­structed, some­body writes the con­struc­tion rules. For most of the numbers that matter, that some­body is a small organ­i­sa­tion almost nobody outside finance has heard of, working from a method­ol­ogy doc­u­ment almost nobody has read.

Modern finance runs on ref­er­ence numbers. A mort­gage resets against a rate. A pension fund owns a slice of an index and is judged against it. A mine sells its gold at a bench­mark. A futures con­tract settles against a ref­er­ence rate. A lending pro­to­col on a blockchain liq­ui­dates a bor­rower when an oracle says the col­lat­eral has fallen far enough. None of these con­tracts names a trade. Each names a number, and a pro­ce­dure for pro­duc­ing it.

Once a number decides who pays whom, the pro­ce­dure stops being a tech­ni­cal­ity. Somebody spec­i­fies the inputs. Somebody spec­i­fies the rule. Somebody runs it and pub­lishes. Every one of those roles creates an incen­tive, and every incen­tive a failure

It was not enough to save the number. The unse­cured inter­bank lending LIBOR was sup­posed to describe had shrunk to the point where most sub­mis­sions could not be tied to real trades. Regulators pushed markets to move to rates built from trans­ac­tions. The US dollar LIBOR panel ended on 30 June 2023.

What replaced it, and what it solves

The replace­ment for dollar LIBOR is SOFR, the Secured Overnight Financing Rate, pub­lished by the Federal Reserve Bank of New York since April 2018. Transactions go in, not answers: the rate on overnight loans secured by US Treasury secu­ri­ties in the repo market, taken as a volume-weighted median across trades that in 2026 have run to roughly three tril­lion dollars a day.

It solves the problem LIBOR died of. A median of that many real trans­ac­tions is very hard to push, and nobody has to guess. It also changed what the number describes. SOFR is overnight and secured against gov­ern­ment bonds, so it con­tains almost no bank credit risk. LIBOR, what­ever its faults, tried to describe what it cost a bank to borrow unse­cured for months. Every bench­mark reports a market, and first it picks which one.

Gold went through its own reform. For almost a century the London gold price was fixed by a small group of banks on a call, in a process dating from 1919. In May 2014 the UK reg­u­la­tor fined Barclays £26 million after a trader manip­u­lated the fix on one day to avoid paying a client. From 20 March 2015 the fix was replaced by the LBMA Gold Price, an elec­tronic auction admin­is­tered by ICE Benchmark Administration, run twice a day at 10:30 am and 3 pm London time, with bids and offers from direct par­tic­i­pants and an audit trail.

Judgement never leaves

Transactions are not the only answer, and in some places they are not an answer at all. The most famous stock index in the world, the S&P 500, is not a list of the five hundred largest American com­pa­nies chosen by formula. Its members are selected by a com­mit­tee at S&P Dow Jones Indices, apply­ing pub­lished eli­gi­bil­ity cri­te­ria and its own dis­cre­tion. Trillions of dollars in index funds buy what­ever that com­mit­tee adds.

Closing prices, the numbers that value most funds each evening, are increas­ingly set in a closing auction, a single match­ing at the end of the day into which a large share of index-fund trading is delib­er­ately con­cen­trated. The rules of that auction, how imbal­ances are pub­lished and when orders may be can­celled, decide the number that becomes the day's offi­cial price.

Liquidity: enough on the other side that none of it hurts.

A real book, mea­sured

Here is the sell side of Binance's bitcoin book 470 mil­lisec­onds before our second, from our own capture. The best offer was $75,807.20 for 6.83704 bitcoin. Behind it, twenty price levels were visible to us, running up to $75,811.26. In total those twenty levels offered $585,236 of bitcoin.

Walk orders of dif­fer­ent sizes up that book, as if they had arrived at that instant, and the answers are on page 40. An order for a thou­sand dollars, ten thou­sand or a hundred thou­sand fills entirely at the best offer. The first price on the screen was the true price for all three.

A million-dollar order does not fit. The whole visible book covers $585,236. That part would have filled at an average only 0.04 basis points above the screen, about $2 in total, which is a remark­ably small toll for more than half a million dollars of bitcoin. The other $414,764 we cannot price at all, because our capture only sees twenty levels. More sellers almost cer­tainly stood above. How many, and where, we can't say.

Half a second later the same book looked very dif­fer­ent. At 537 mil­lisec­onds into the second, after a run of buying on every venue, the twenty visible levels on the sell side held $30,094. A hundred-thousand-dollar order, which had filled at one price a moment earlier, now could not be priced from what was visible at all.

How much does a million dollars cost?

Odd ques­tion, on purpose. Not what a million buys: what it costs to spend. The honest answer has three parts, and only the first ever reaches a screen.

The first is the visible cost: fees, and the spread you cross. On Binance at our instant the spread was one cent, so cross­ing it cost nothing worth mea­sur­ing. Fees are pub­lished, and depend on who you are.

The second is slip­page against the visible book, which is what we cal­cu­lated above, and which can be tiny or large depend­ing on the moment you arrive.

The third is what happens next. Other par­tic­i­pants see the trades. Market makers who were offer­ing size move their prices away, because a large buyer is infor­ma­tion. Some pull their offers entirely. The stair­case moves when you step on it, so large orders are almost never sent at once. They are cut into small pieces and released over minutes or hours by algo­rithms with names like VWAP and TWAP, which try to buy at the average price of the period without announc­ing that a big buyer is present.

Slicing costs too: while you wait, the price wanders for its own reasons. Every exe­cu­tion desk trades moving the

market against waiting for it, and the screen shows neither.

The queue

Each step of the stair­case is also a queue. Most exchanges fill resting orders at the same price in the order they arrived, which traders call price-time pri­or­ity. A seller who joined the best offer early is filled first; one who joined later waits, and may not be filled at all before the price moves away.

That makes the book a strange kind of object. The 6.83704 bitcoin offered at $75,807.20 in our snap­shot was not one seller's offer. It was a line of them, and the posi­tion of each in the line was worth some­thing. Firms pay heavily for the speed to reach the front. The front is filled by traders cross­ing the spread; the back is filled mostly just before the price turns against it.

The same arith­metic, selling

Everything above applies in reverse to a seller, with one asym­me­try that matters in bad weeks. When a large holder must sell, because of a margin call or a redemp­tion, the buyers in the book know that the selling has to happen. They have no reason to stand in its way. Bids thin out, and the seller walks further down the book than a buyer of the same size would have walked up it. Forced sellers pay for being forced, and the price they get is one they helped to make.

So the value of a port­fo­lio on a screen, quan­tity times last price, is money no large holder could actu­ally receive. It assumes selling one unit at a time into a market that never notices. Close enough for a small investor. For a fund that owns a real share of what it holds, the gap is one of its largest risks, and the screen never shows it.

The part you cannot see

Order books also contain orders that do not show. Many venues allow iceberg orders, which display a small visible size and replen­ish from a hidden reserve as they fill. Some trading happens away from public books entirely, in dark pools and through dealers quoting pri­vately. A visible book is a lower bound on liq­uid­ity, and on a bad day it is an opti­mistic one, because the orders it shows can be can­celled faster than a human can click.

So a large buyer is really asking what the price will be after they have fin­ished. That number does not exist yet, and their own order is one of the things that will decide it.

The next four pages show one real order that was large enough to test the book, and what the book did around it.

Gold is gold. An ounce of it is 31.1035 grams of the same element wher­ever it is. And yet on any given morning there are a dozen prices for an ounce of gold, all of them pub­lished, all of them correct, and no two quite the same.

There is the London price, for metal held in the vaults of the London bullion market. There is the price of a gold future in New York, for a hundred ounces deliv­ered to an approved depos­i­tory in a named month. There is the price of a share in a gold fund, a coin at a dealer, a token on a blockchain, and gold in Shanghai, quoted in yuan per gram. The natural reac­tion is to see arbi­trage every­where: surely some­body should buy the cheap one and sell the dear one until they meet.

Sometimes some­body does. Mostly they were never meant to meet. The metal is the same; the con­tract around it differs, and a price is always the price of a con­tract.

Gold with a pass­port

Location can become a price on its own. The Shanghai Gold Exchange quotes gold in yuan, and China con­trols how much gold can be imported. When domes­tic demand runs hot and imports cannot keep up, Shanghai gold trades at a premium to London con­verted at the exchange rate, some­times a large one. No arbi­trageur will erase it by lunch. It is the price of the border.

Coins at a dealer sit at the other end. A one-ounce bullion coin sells above the value of its metal, because some­body had to mint it, ship it, insure it and hold it in a shop, and because a dealer buys it back from you below the price they sell it at. Small buyers pay for small­ness.

Time has a price in gold too

The dif­fer­ence between a gold future and gold today is usually small and pos­i­tive, and it has a name: the cost of carry. Somebody who buys metal now and sells a future against it must finance the pur­chase and pay for storage and insur­ance until deliv­ery. The future is priced so that doing so earns roughly nothing extra. When inter­est rates rise, the gap between spot and futures widens, because car­ry­ing the metal costs more. When it inverts and futures trade below spot, some­body is paying a great deal to have gold now rather than later, which is itself a piece of news.

A token has its own version. PAX Gold and Tether Gold can trade con­tin­u­ously, includ­ing at week­ends when the London market is shut, so for two days a week their prices are among the few live gold prices in the world. They are also, during those hours, prices with no market in the metal behind them to arbi­trage against. Their weekend prices are a view about Monday.

Not every gap is free money

None of this means the prices of related instru­ments drift wher­ever they like. The links are real and they are enforced by people who make their living from them. If a gold fund trades far enough above the value of its metal, autho­rised par­tic­i­pants create new shares by deliv­er­ing gold and sell them. If a future trades too far above spot plus the cost of carry, someone buys the metal, sells the future and deliv­ers. The gaps close to roughly the cost of doing that, and no further.

That resid­ual gap is the most infor­ma­tive number of all. It is the price, set by a market, of the spe­cific rights and fric­tions that sep­a­rate one kind of ounce from another.

An ounce is an ounce. The con­tracts around it never are, and the price tells you so, if you read what it is the price of.

Two screens show the same price. One market can absorb a large order without flinch­ing. The other would move several per cent. Nothing on either screen tells you which is which.

That is the dif­fi­culty with liq­uid­ity. It is the most impor­tant prop­erty of a market for anybody who trades size, and it has no ticker symbol. It lives below the dis­played quote, in how much is offered at each price behind the best one, how quickly those offers come back after they are taken, and how many of them are real. Most of the time it is invis­i­ble because it is plen­ti­ful. It becomes visible at the moment it dis­ap­pears.

Same price, dif­fer­ent market

Our own capture con­tains a clean example, one second apart on the same venue. At 537 mil­lisec­onds into our second, Binance's best offer for bitcoin was $75,844.01. At 1,597 mil­lisec­onds, just after the second ended, it was $75,840.34. The two prices are $3.67 apart, half a basis point, close

enough that nobody glanc­ing at a screen would notice any change.

The markets under­neath were not close. At +537 ms the twenty visible levels on the sell side held $30,094. At +1,597 ms they held $638,838, about 21 times as much.

Send the same order into both. A $25,000 pur­chase against the first book would have used 14 of the twenty levels and paid 0.11 bp over the screen. Against the second book it would have filled at the best offer. A $250,000 pur­chase could not have been priced at all from the first book, and would have cost next to nothing against the second. Same price. Different market.

Where it went

The thin book had a cause. It sat in the middle of a burst of buying that ran across all five venues, after the large order on page 44. Market makers who had been offer­ing size either sold it or moved away. Within about a second they were back, at slightly higher prices, deeper than before. On a calm day that cycle happens con­stantly and nobody sees it, because the price barely moves.

Spreads tell a similar story. At the start of the second, the gap between best bid and best ask was a cent on Binance and Coinbase and ten cents on Bybit, Gate and OKX. By the end of the second the gap on Coinbase was $9.37 and on Gate $3.60. On the other three it had not moved. Nobody announced it. On two venues the people willing to act stepped back, and the price stayed on the screen as though nothing had hap­pened.

When it goes all at once

Those were small, ordi­nary with­drawals, repaired within a second. Larger ones are the moments market history remem­bers. On 6 May 2010 US equity markets fell and recov­ered within about half an hour, and some shares traded at a cent or at $100,000 as market makers' place­holder quotes, known as stub quotes, sud­denly became the only orders left in the book. On 15 January 2015, when the Swiss National Bank stopped holding the franc below 1.20 per euro, liq­uid­ity in one of the world's most traded cur­rency pairs van­ished for minutes and prices gapped by double digits in per­cent­age terms.

In each case the dis­played price before the event said nothing about what was about to happen, because a dis­played price describes the best offer and says nothing about the ones behind it. Depth that has been there every day for years prom­ises nothing. The orders that make it up are free to cancel, and they cancel together, because the people who place them are react­ing to the same infor­ma­tion at the same time.

Why the makers leave together

It is tempt­ing to see the with­drawal of market makers in a fast market as deser­tion. From the inside it looks like arith­metic. A market maker earns the spread on trades with people who have no special infor­ma­tion, and loses on trades with people who know some­thing she does not. In calm markets the first kind dom­i­nates. When prices start moving fast, a larger share of the orders arriv­ing are from people react­ing to news she has not yet seen. Every quote she leaves in the book becomes an option she has given away for free.

So she widens her spread, reduces her size, or leaves. So does every other maker running a similar model on similar data, at the same moment. No single maker ever promised the liq­uid­ity that had been there for years. It was the sum of many sep­a­rate deci­sions to stay, and the same signal reverses all of them.

That is also why it comes back. The first makers to return to a thin book are paid well, in wide spreads, for taking the risk. Within our frozen second, the Binance book went from thin to deeper than it had started in about one second. On the days market history remem­bers, the same process took minutes, and in a few cases exchanges had to stop trading to give it time.

Reading what cannot be seen

Professional traders there­fore watch things the price does not show: how much is offered within a few basis points of the best price, how that changes through the day, how quickly the book refills after it is hit, how wide spreads are rel­a­tive to their normal range. None of them is a fore­cast. Each is a way of asking the market how much weight its current price could bear.

They also know that the most dan­ger­ous moment is not a thin book but a thick one that nobody has tested. Page 42 shows Binance's visible sell side over the minute around our second. Three seconds before it, more than a million and a half dollars was on offer. Inside the second, thirty thou­sand. Nothing about the dis­played price warned of either number.

Same number, one second apart. Behind the second one stood many times the money, and the screen never said a word about it.

On 16 September at 15:38:52.188 UTC we asked nine public price ser­vices the same ques­tion at the same moment: what does one bitcoin cost? Every request left our server within a few mil­lisec­onds of the others. All nine answered. No two gave the same number.

The answers ran from $75,666.93 to $75,742.70, a range of $75.77, about 10 basis points on a single asset, asked a single ques­tion, in the same second. It was a dif­fer­ent second from the one the rest of the issue is built on: price ser­vices cannot be asked about the past, so we asked them about a present of our own, and every page that shows these answers is stamped with that other second.

None of them is wrong. Each answered a slightly dif­fer­ent ques­tion and says which one in its doc­u­men­ta­tion. The dis­per­sion is the visible dif­fer­ence between def­i­ni­tions.

Nine ques­tions that sound like one

Three ser­vices, Kraken, Bitstamp and Gemini, returned the last trade on their own books: the price at which some­body most recently bought or sold there, however small the trade and however long ago. Coinbase returned its own spot price for its dollar market. OKX, Deribit and Bybit returned indices, each built from a basket of other exchanges by their own rules, the kind of number page 19 took apart. CoinGecko returned an aggre­gate across a very large number of venues.

Binance returned some­thing else again: the average trade price over the pre­vi­ous five minutes, in tether. It answered a ques­tion about the recent past, not the present, in a dif­fer­ent unit. Bybit's index was also in tether. The two highest answers in our sample were the two denom­i­nated in tether, which is exactly what page 12 would predict.

The ages

Some ser­vices stamp their answers with the time the number was made. Most of those stamps were a frac­tion of a second old. CoinGecko's was 112 seconds old when it reached us. It was not broken; it refreshes on a sched­ule, and between refreshes the last number stands, pre­cisely like a quote on a quiet venue. Three ser­vices gave no time­stamp at all, so we cannot say how old their numbers were. We can only say when we received them.

Which one was right?

Each, for its purpose. A trader on Kraken needs Kraken's book. A futures exchange decid­ing whether to liq­ui­date some­body needs an index no single venue can push. An accoun­tant needs a doc­u­mented source at a doc­u­mented time. The mistake is the belief that there was a tenth number under­neath the nine that was simply the price. If machines looking at the most liquid object of its kind dis­agree by tens of dollars, nobody should be sur­prised when people pricing a paint­ing, a house or a used car dis­agree by far more.

We asked nine times and got nine answers. We report all nine, with their def­i­ni­tions, their ages and their units, and we do not average them.

We see faces in clouds

The human habit of finding mean­ing­ful pat­terns in noise has a name, apophe­nia, and in markets it has a long paper trail. In 1959 the statis­ti­cian Harry Roberts pub­lished a short article in the Journal of Finance showing that series built from random numbers produce the same head-and-shoulders for­ma­tions and support levels that chart readers were selling as signals. The article did not stop anyone. The pat­terns are too per­sua­sive to give up because a statis­ti­cian gen­er­ated some.

The science of streaks is humbler than either side of the argu­ment would like. In 1985 Thomas Gilovich, Robert Vallone and Amos Tversky ana­lysed bas­ket­ball shoot­ing and con­cluded that the hot hand, the belief that a player who has hit several shots is more likely to hit the next, was an illu­sion. Three decades later Joshua Miller and Adam Sanjurjo showed that the orig­i­nal method con­tained a subtle sta­tis­ti­cal bias, and that cor­rected, the data suggest a modest hot hand after all. Streaks are neither always real nor always false, and careful experts misread them, in both direc­tions, for thirty years.

Sometimes some­body does know some­thing

The feeling that a move means some­thing is far from stupid. Sometimes it is exactly right. Markets contain people who know more than others: about a company, about a large order that is coming, about a reg­u­la­tor's deci­sion. Economists model it directly. In Albert Kyle's 1985 model an informed trader hides their buying inside ordi­nary noisy flow, and a market maker, unable to tell which orders are informed, moves prices against all of them a little.

That is what page 55 of this issue shows from the other side. Nineteen mil­lisec­onds after a burst of buying on Bybit, the spread there went from ten cents to $12.80. Whoever was quoting did not know whether the buyer knew some­thing. They behaved as if they might. Market makers do not need to find the informed trader. They only need to price the pos­si­bil­ity, and that pos­si­bil­ity is part of every spread you cross.

Somebody usually does know some­thing. The trouble is that the price alone cannot tell you which moves carry infor­ma­tion and which are the coin landing heads four times in a row.

Story, then flow, then price

Most of the time what moves a price is a story. The econ­o­mist Robert Shiller has argued that nar­ra­tives spread through a pop­u­la­tion much as epi­demics do, and that their spread changes what people do with money. A story becomes flow when enough people act on it. Flow becomes price when

there is not enough on the other side of the book.

In January 2021 a story about a heavily shorted video game retailer spread through an online forum. Shares of GameStop, which began the month under $20, traded as high as $483 on 28 January. That morning several retail brokers restricted cus­tomers from buying more. The price fell by more than half within days. Nothing about the company's shops had changed during those weeks. The story, the flow and the book had.

The number you saw first

Patterns are one trap. The price we already know is another. In a 1974 exper­i­ment Amos Tversky and Daniel Kahneman spun a wheel of fortune, rigged to stop at 10 or at 65, in front of their sub­jects, then asked what share of African coun­tries were members of the United Nations. The wheel had nothing to do with the ques­tion. Those who saw 10 gave a median answer of 25%. Those who saw 65 said 45.

Markets are full of wheels. The price you bought at, last year's high, a round number on a screen: each one pulls the next judge­ment towards itself. In 1998 Terrance Odean fol­lowed about 10,000 accounts at a dis­count broker and found that investors sold shares that had risen since pur­chase far more readily than shares that had fallen. The winners they sold went on to beat the losers they kept by 3.4 per­cent­age points over the fol­low­ing year. The pur­chase price was infor­ma­tion about the investor. They treated it as infor­ma­tion about the share.

Page 11 asks five ques­tions of every number, and one of them is how old it is. The price in your head is the oldest number in the room, and the only one that feels like yours.

What the pat­terns cost

Acting on pat­terns is expen­sive in a way that is easy to measure. In a study of more than 66,000 American bro­ker­age house­holds from 1991 to 1996, Brad Barber and Terrance Odean found that the fifth who traded most earned an average net return of 11.4% a year, while the market returned 17.9. Their trades were not all wrong. They were numer­ous, and each one paid a spread and a com­mis­sion, the same arith­metic that page 40 of this issue shows for a small order in a thin book.

Nobody can stop seeing pat­terns. The defence is to ask, before acting on one, what it would look like if it were noise, and to remem­ber that page 92 is the answer to that ques­tion.

A price is where stories meet money. The chart records the meeting. It does not explain it.

Every gov­ern­ment that decides whether to build a safer road, tighten an air pol­lu­tion rule or require a new kind of brake has to compare a cost in money with a benefit mea­sured in deaths avoided. To compare them it needs a number.

In the United States that number is called the value of a sta­tis­ti­cal life. The Department of Transportation's guid­ance set it at $13.2 million for analy­ses in 2023 dollars and $13.7 million for 2024. The Environmental Protection Agency uses its own figure, updated for infla­tion and income. Other coun­tries use other numbers and other names. Britain's Department for Transport speaks of the value of pre­vent­ing a fatal­ity.

The phrase invites a mis­un­der­stand­ing worth taking apart slowly. Nobody is pricing one person's life. No agency believes a person could be bought for $13.7 million, or that a life is worth that much and not a dollar more. The number refers to no iden­ti­fi­able person at all.

What the number actu­ally mea­sures

Start from small risks. Suppose a reg­u­la­tion would reduce the annual risk of dying in a par­tic­u­lar kind of acci­dent by one in 100,000, for each of a million people. Nobody knows who would have died. Across the million, the reg­u­la­tion is expected to prevent 10 deaths a year. Those are the sta­tis­ti­cal lives.

Now ask what people are willing to pay for small reduc­tions in their own risk of death. Economists esti­mate this from how much extra pay workers accept for riskier jobs, how much people pay for safer prod­ucts, and from surveys. If people, on average, would pay about $137 a year to reduce their own annual risk by one in 100,000, then a million people together would pay $137 million for a change that saves ten sta­tis­ti­cal lives. Divide one by the other and you get $13.7 million per sta­tis­ti­cal life.

That is the whole con­struc­tion. The value of a sta­tis­ti­cal life is a way of adding up many people's val­u­a­tions of tiny changes in risk. It

is a price for risk, expressed per expected death, used so that a benefit mea­sured in safety can be set beside a cost mea­sured in money.

Why it exists

Refusing to use a number does not avoid the com­par­i­son. Every deci­sion to spend or not spend on safety makes one implic­itly. A rule that costs $500 million and is expected to prevent one death a year for twenty years has an implied value per sta­tis­ti­cal life whether anyone writes it down or not. Writing it down makes the com­par­i­son con­sis­tent across agen­cies and visible to the public, which can then argue with it.

It is also con­tested, and the con­tests are serious. Should the number differ by income, age or country? Agencies in the United States gen­er­ally use one value for every­one, pre­cisely to avoid valuing a poorer or older person's risk reduc­tion less. International com­par­isons that adjust for income produce lower numbers for poorer coun­tries, which is defen­si­ble as a measure of will­ing­ness to pay and uncom­fort­able as a state­ment about people.

A rule that failed the test

In 2002 a two-year-old named Cameron Gulbransen was killed when his father reversed the family car in their drive­way and could not see him. Congress passed a law named after him in 2008 that told the Department of Transportation to improve rear vis­i­bil­ity. The rule, pub­lished in 2014, required rear-view cameras in new light vehi­cles from May 2018. The agency's own anal­y­sis put the rule's own effect at 13 to 15 deaths pre­vented a year, at a net cost of $15.9 million to $26.3 million for each equiv­a­lent life saved: well above the value of a sta­tis­ti­cal life it used at the time. Once every vehicle on the road carried a camera, it expected 58 to 69 lives saved a year. The rule went ahead because Congress had required it. The gap was written down, pub­lished, and chosen in the open.

Values also differ between agen­cies and between coun­tries. Britain's trans­port appraisal uses a figure far lower in pounds than the American one in dollars. The spread is no scandal. This issue found the same thing in nine price ser­vices answer­ing one ques­tion: dif­fer­ent def­i­ni­tions, dif­fer­ent inputs, dif­fer­ent pur­poses.

The bound­ary of the issue

This is as far as the idea of a price can be pushed. A value of a sta­tis­ti­cal life is con­structed, like an index. It is con­tested, like a bench­mark. It depends on who is asked and how, like an appraisal. And it is used to decide real things, like a mark. Everything this issue has said about prices applies to it.

What does not apply is exchange. Nobody can sell a sta­tis­ti­cal life or buy one. The number exists because deci­sions must be made and money must be com­pared with some­thing that is not money. It is a price in every sense except the one most people mean.

The number was never about one person. It is about every­one, a little.

The number at the top of this page was on our screen at 07:24:04 on 16 September, in the same second this whole issue is built on. It was Kraken's best bid for bitcoin, as our col­lec­tor recorded it: $76,007.30, for 0.0537 bitcoin.

It was $256 above Kraken's own best offer at the same moment, which is impos­si­ble in a working order book: a buyer willing to pay more than a seller is asking would simply have traded with them. The level had last traded 18 minutes earlier, at 07:06:00. It stayed at the top of our record of Kraken's book, unchanged, for 56 minutes.

The venue had not made a mistake. We had. Our col­lec­tor sub­scribed to the top ten levels of Kraken's book. When a level moved out of that window, Kraken simply stopped sending updates about it, and our code, which only removed levels when told their size was zero, kept it forever. We found the bug while pre­par­ing this issue, fixed it the same day, and with­held Kraken from every table in the issue. But the ghost bid is a perfect spec­i­men for this article, because it looked exactly like a price.

Is it still a price?

Every dis­played price is a state­ment about the past. For a liquid market the past is mil­lisec­onds old and the state­ment is nearly true of the present. As the age grows, the same number means less. It does not become wrong all at once. It decays.

A last trade from a second ago tells you roughly where you could trade now. From an hour ago, it tells you where the market was. From a month ago, it tells you what some­body once paid. From five years ago, on an instru­ment that has not traded since, it tells you almost nothing about value and a good deal about history.

Screens rarely say so. Most dis­plays show a number without its age, in the same type and colour whether it is fresh or years old.

How prices go stale

Trading halts are the every­day case. An exchange stops trading in a share pending

news, and the last price sits on every screen until trading resumes. It may resume far away. Circuit break­ers do the same across whole markets for minutes.

Closed markets are the larger case. After Russia invaded Ukraine in February 2022, the Moscow Exchange kept its stock market shut from 28 February until 24 March, and the London Stock Exchange sus­pended trading in the deposi­tary receipts of many Russian com­pa­nies in early March. For weeks their last prices remained on screens and in index cal­cu­la­tions and fund val­u­a­tions, describ­ing a market that did not exist in any form an investor could use.

Delisting leaves a last price per­ma­nently. A company taken private, a fund wound up, a token whose venues have closed: each has a final print that will stay in data­bases forever, per­fectly accu­rate as history.

Private assets live per­ma­nently in this con­di­tion. A private equity fund reports the value of com­pa­nies that have not traded, usually quar­terly and with a lag, based on appraisals, com­pa­ra­ble listed com­pa­nies and recent funding rounds. Its reported value can stay calm through a market fall that would have repriced the same com­pa­nies on an exchange within days.

The quiet version

Most stale prices are not dra­matic. A bond that trades a few times a month is marked every day by a pricing service using quotes and models. A prop­erty fund values its build­ings every quarter. A small company's shares trade a few hundred times a day, and its last price at the close may be from a trade of a few shares an hour earlier. Each of these numbers is rea­son­able. Each is also, quietly, a state­ment about some time ago, set in the present tense.

The danger lies in com­bin­ing them. A port­fo­lio that holds a liquid index fund and an illiq­uid private fund shows a total that adds a price from this minute to a price from last quarter. In a falling market the total looks stead­ier than any­thing inside it really is.

What to ask of any number

The defence is simple and almost never used. For any price, ask how old it is. Ask when the instru­ment last traded, and in what size. Ask whether the market it came from is open. Ask whether the source is still sending updates, or whether a number is merely still there, which, as our own col­lec­tor demon­strated, can look iden­ti­cal.

Our ghost bid was a real number, cor­rectly stored, faith­fully dis­played. It had simply stopped being a price 56 minutes before we noticed, and nothing on the screen said so.