EW JOURNAL

Six impossible things before breakfast is the boast the White Queen makes to Alice in Through the Looking-Glass, offered as a kind of daily exercise, the way someone else might mention a morning jog. Pressed by Alice, who insists that one cannot simply believe impossible things, the Queen explains that practice is the whole secret: “I daresay you haven’t had much practice. When I was your age, I always did it for half an hour a day.” Lewis Carroll wrote that exchange under a pen name. Under his own name, Charles Dodgson held the mathematical lectureship at Christ Church, Oxford, for twenty-six years, from 1855 to 1881, publishing Euclid and His Modern Rivals, two volumes of Symbolic Logic, and a study of parliamentary voting methods still cited today as Dodgson’s method. The second volume of Symbolic Logic introduced what later logicians recognised as the earliest printed use of a truth tree, a method for checking whether a whole set of propositions can be true together or are quietly doomed to contradict one another several steps further down the page. A man who spent his career testing whether statements could survive each other’s company had a professional head start on a much harder sorting problem: telling a thing that cannot happen apart from a thing that merely almost never does. Financial risk hides almost entirely inside that second category, the one perpetually mistaken for the first.

Human intuition keeps a single working category for anything sufficiently rare: it is simply not going to happen, to anyone, this week. Set a one-in-ten-thousand chance beside a flat zero, and intuition cannot tell them apart: both collapse into that same category almost instantly, because intuition evolved to weigh a few dozen daily decisions, not to track a number past the point where a human nervous system can still feel the difference between faint and absent. A lottery ticket and a long-haul flight carry odds separated by several orders of magnitude, one a near-impossibility and the other a near-certainty of safe arrival, and yet most people spend more active worry on the flight than they ever spend hoping about the ticket, because dread and hope answer to imagination long before they answer to arithmetic, and a vivid enough story beats an accurate enough number every time the two compete for the same attention.

Arithmetic keeps no such category and does not soften for how a number feels. Multiply a one-in-ten-thousand chance across ten thousand independent tries, and the mathematics expects the event to appear at least once more often than not; failing to see it even once across that many tries would be the actual surprise. A single try is the only thing rarity ever actually measures. Stretch the number of tries far enough, and that same measurement stops predicting anything useful about the outcome at all. Statisticians sometimes describe the general pattern as the law of truly large numbers: given enough independent trials, almost any event, however improbable on any single attempt, eventually becomes close to certain.

Such institutions are, functionally, machines built to generate independent tries at enormous scale: a trading book with thousands of live positions, a clearing operation processing millions of transactions a year, decades of continuous operation layered on top of both. Each position or transaction is its own small room, carrying its own small, genuinely low chance of catching fire, and a single room by itself rarely troubles anybody. Enough such rooms held for long enough is an entirely different calculation, one in which the odds quietly compound until some room among the thousands is, for all practical purposes, guaranteed to catch.

Tallying is a risk committee’s real work, counting how many small, individually reasonable risks the institution is currently running at once, and checking whether any of them are secretly correlated, wired, however indirectly, to the same underlying event. Almost every risk anybody ever actually takes looks small enough to ignore, judged one at a time. That is precisely why it gets taken, and precisely why counting matters so much more than judging any single instance ever could on its own.

The sociologist Charles Perrow gave this structural inevitability a formal name in 1984, in a book called Normal Accidents: Living with High-Risk Technologies, written partly in response to the partial reactor meltdown at Three Mile Island in Pennsylvania on 28 March 1979. Perrow’s actual claim was narrower and more interesting than it sounds: in any system combining genuine complexity, too many interacting parts for a single operator to track in full, with tight coupling, too little slack for a developing problem to be caught and isolated before it cascades into the next part of the system, accidents stop being occasional surprises and become a predictable feature of the system’s own structure. Three Mile Island supplied the founding case. A relief valve lifted correctly when reactor pressure climbed, exactly as its design required, and then failed to reseat once that pressure had fallen back; coolant kept draining through it for more than two hours while a light on the control panel reassured the operators watching it. That light reported only the power state of the solenoid commanding the valve, a fact entirely true on its own terms, and operators who had never been trained to see the gap took it as proof the valve itself had closed, a fact that happened to be false. Dozens of alarms competed for the same few seconds of attention, none of them ranked by importance, and a true reading feeding a false conclusion is precisely the combination nobody on shift that night had been trained to catch. Perrow had, by his own later account, never given industrial accidents a moment’s serious thought before being invited onto the government’s own investigative effort afterward, which is itself a small illustration of how invisible a risk can remain right up until the one occasion it stops being invisible.

Perrow himself eventually extended the diagnosis to finance directly, telling the economics writer Tim Harford in 2011 that modern finance was a perfect example of a complex, tightly coupled system whose complexity exceeds the complexity of any nuclear plant he had ever studied. The 2008 collapse of AIG supplied as clean an illustration as Three Mile Island had. AIG Financial Products, a single subsidiary run out of Wilton, Connecticut, with a trading desk in London doing much of the actual work, had added more than five billion dollars to AIG’s own pre-tax income between 1987 and 2004, a run of profit so dependable that the parent company’s market value grew from eleven billion dollars to a hundred and eighty-one billion dollars across those same years. The unit had sold credit default swaps, insurance-like contracts that pay out if an underlying bond defaults, across a book that regulators and the firm’s own disclosures put at somewhere around four hundred and forty billion dollars in the highest-rated mortgage tranches alone. Joseph Cassano, the division’s head, told investors in August 2007 that it was hard for anyone there to even see a scenario, within any kind of realm of reason, that would see the firm losing one dollar on any of those transactions. The remark reads today as the purest possible statement of the single-room fallacy: confident, specific, and true of almost every individual contract taken entirely on its own.

Those contracts had been priced and sold as though each one were its own small, independent room. The 2008 housing collapse showed they shared a single floor: when American mortgages failed, they failed together, in the same months, for the same underlying reason, asking AIG to pay out on an enormous number of its rooms all at once, nothing like the small, manageable trickle its own models had actually planned for. The federal government committed roughly a hundred and eighty-two billion dollars to keep the company from collapsing entirely, eventually recovering slightly more than it had committed, a rare happy ending to an otherwise textbook demonstration of the same mechanism Perrow had already described, in a different industry entirely, three decades earlier.

A thousand boxes can stand in a row, each one correctly labelled with a vanishingly small chance of catching fire, and the label on every single box can be entirely true. None of those labels says anything about whether the boxes happen to be stacked against the same wall, wired to the same faulty circuit, or warmed by the same failing furnace somewhere underneath all of them. Asking whether any one box is likely to burn is the easy, comfortable question, and it is also the wrong one. Asking how many boxes just like it are quietly sharing a wall is the question that actually decides what happens to the building. Dodgson’s own truth tree tested whether a set of propositions could all hold at once without quietly contradicting one another somewhere further down the page; a risk committee is really running the same test on a portfolio instead of a syllogism, and the single occasion it fails is the only one that ever gets remembered afterward.