Twenty Good Years

In the year 2000, Volkswagen delivered about 331,000 cars in China. By 2018 it was delivering 4.2 million. Its Chinese joint ventures alone threw off €5.2 billion of operating profit in 2015 — a single country supplying, by analysts' reckoning in the best years, something close to half of the group's earnings. Twenty years is a long time for a line to behave itself. Long enough that it stops looking like a line and starts looking like a fact about the world.

In 2024, those same joint ventures returned €1.7 billion — a third of what they had managed nine years earlier. Deliveries fell to 2.93 million, and to roughly 2.69 million in 2025. Volkswagen's share of the Chinese market slid from 19.3% in 2020 to 14.5% in 2023. In 2023, after some fifteen years at the top, it stopped being China's best-selling car brand. BYD was.

It is tempting to file the rest under hubris, and there is a great deal to file. A corporate airline. A software subsidiary that could not ship. Consultants. A chief executive paid €10,345,677 in 2024. Every one of those is real, and every one of them is a story told afterwards — which is the trouble with it. Hubris would have explained the previous twenty years just as fluently, if the line had kept going up. Then it would have been called conviction.

So put the morality down and ask the narrower, more uncomfortable question. At what point should Volkswagen have stopped believing the line? Not: when should it have been humbler. When, specifically, in which year, on the strength of which number, should a reasonable person looking at twenty years of Chinese growth have written a different plan? The honest answer — the one that ought to worry you about your own numbers — is that the line was never going to tell them. It could not. That is not a fact about Volkswagen. It is a fact about lines.

The oldest unsolved problem in forecasting

David Hume put it in 1748, and nobody has answered him since. The past is evidence about the past. Turning it into a claim about the future requires an extra premise — that the future will resemble it — and that premise cannot itself be established by observing that it has held so far, because that is the very move under suspicion.

"The bread, which I formerly eat, nourished me… but does it follow, that other bread must also nourish me at another time? The consequence seems nowise necessary."

Bertrand Russell, in 1912, gave the problem the animal it deserved. "The man who has fed the chicken every day throughout its life at last wrings its neck instead," he wrote, "showing that more refined views as to the uniformity of nature would have been useful to the chicken." His point was not that the chicken reasoned badly. The chicken reasoned exactly as we do about sunrise. "Our instincts certainly cause us to believe that the sun will rise to-morrow," Russell went on, "but we may be in no better a position than the chicken which unexpectedly has its neck wrung."

Nassim Taleb kept the bird and changed the calendar. His turkey is fed every day by people who appear to have its interests at heart, and "every single feeding will firm up the bird's belief that it is the general rule of life to be fed every day by friendly members of the human race." Then comes the Wednesday before Thanksgiving. The sentence that matters is the one after, because it is not about poultry, and it is not a joke:

"Its confidence increased as the number of friendly feedings grew, and it felt increasingly safe even though the slaughter was more and more imminent. Consider that the feeling of safety reached its maximum when the risk was at the highest!"

Read that as a claim about measurement and it stops being philosophy. The turkey's confidence and the turkey's danger were not merely both rising. They were rising for the same reason, out of the same data, at the same time. Every additional feeding was simultaneously another point of evidence and another day closer. Which raises a question you can actually settle with arithmetic: when you build a proper statistical model of a trend, does its confidence behave like the turkey's?

It does. Provably, and worse than you would like.

Why the error bars point the wrong way

Fit a straight line to a growing series — the humblest thing a forecaster can do. The uncertainty the model reports around its own slope shrinks with the square root of the number of observations, and shrinks further the more tightly the points hug the line. The band it draws around a future value behaves the same way. This is not a flaw in the method. It is the method working correctly. Two consequences follow, and they are the whole essay:

• Every year the trend survives, the forecast gets narrower. The model's certainty is a direct function of how long the world has been consistent — which is precisely the quantity the turkey was tracking.
• The calmer the history, the faster it narrows. A smooth, well-behaved, boringly reliable series produces the most confident predictions. A jumpy one produces humble ones. The safest-looking business is the most dangerous to extrapolate.

Now the part that closes the escape hatch. You might think: fine, so widen the bands. But you cannot know to. Before the break, the model is not overconfident — it is correctly calibrated. Give it a real walk-forward backtest, refitting each year and scoring it on the year it has not yet seen, and it passes. In the simulation below, run across a couple of hundred synthetic histories, about 95% of years land inside the 95% band, exactly as advertised. The model has an honest résumé. It earns it. It keeps earning it right up to the last good year, and the résumé is worth nothing, because it was compiled entirely inside a world that is about to stop existing.

A confidence interval measures how consistent the past has been. It has never once been a statement about the future.

Run the simulation below and you can watch where the confidence peaks. At its default settings it peaks, about half the time, on the last good year — the turkey's Tuesday, drawn as a bar chart. That is not luck. Certainty here is manufactured out of consecutive quiet years, so it is necessarily highest on the last of them. And if you turn the noise up, something worse happens: the peak wanders past the break, because a scattered model needs several bad years before they can outvote the calm ones. It does not merely fail to see the break coming. For a while, it fails to notice it arriving.

Twenty-five standard deviations

On August 13, 2007, with one of its funds down some 27% on the year, Goldman Sachs's chief financial officer, David Viniar, explained the situation to the Financial Times. "We were seeing things that were 25-standard deviation moves," he said, "several days in a row."

Four researchers took him at his word and did the arithmetic. Under the bell curve his models assumed, a 5-sigma daily move should turn up about once every 13,932 years. A 6-sigma move, once every 4 million years — roughly the age of our species. An 8-sigma move, once every 6.4 trillion years, comfortably longer than the universe has existed. And Viniar's 25-sigma day should arrive about once every 1.3 × 10¹³⁵ years. He had seen several in a row. Two of them back to back, the authors reckon, is about as likely as winning the national lottery 42 times running.

"Would you prefer the people looking after your money to be incredibly unlucky, or just plain incompetent?"

There is an obvious rescue, and it is worth watching it fail. The rescue is fat tails: real markets throw up extremes far more often than a bell curve allows, so the true odds were never that long. Quite so — and, the authors note, beside the point. "So what if the true probability of a 25-sigma event is 3.057e-36 rather than 3.057e-136?" they ask. "Moving the decimal point a hundred places or more makes no practical difference." Either way the model had ruled the week out of existence. A number that absurd is not a measurement of the market. It is a measurement of the instrument. When your model reports that last week was impossible, it is not telling you about last week.

Volkswagen has its own version of the number, and it is written in its own annual report. In 2018, the company told its shareholders, in plain type: "We are preparing to enable delivery of approximately 400,000 New Energy Vehicles to customers in China in 2020 and approximately 1.5 million in 2025." In 2025, Volkswagen delivered about 116,000 battery-electric cars in China — a figure that had itself fallen 44% in a year. The target counted plug-in hybrids as well, so the comparison is not clean; no honest accounting of it comes anywhere near 1.5 million. That forecast was not a lie, and it was not lazy. It was the confident output of a company reading its own excellent, twenty-year, beautifully consistent line.

What a broken line costs

On the evening of July 9, 2026, Volkswagen's supervisory board sat for hours in Wolfsburg and then published a Zukunftsplan — a future plan, twelve initiatives, a target picture for 2030. What the company confirmed, in its own words, was a program of subtraction: the model range to be streamlined by up to 50%, the configuration and equipment options a customer can order cut by up to 75%. The chief executive's stated reason was weniger Komplexität — less complexity. What the company did not confirm was anything at all about people or places.

Two weeks earlier, Manager Magazin had reported that around 100,000 jobs were at risk and that four German plants — Hannover, Emden, Zwickau, and Audi's Neckarsulm — were slated to close. Those figures have been repeated everywhere since. Volkswagen has confirmed neither a single job nor a single plant. The chair of the group works council, Daniela Cavallo, left the meeting and said that the board's conduct toward the workforce could not be surpassed for disrespect, and gave the chief executive until the following day to say plainly whether the reports were true.

The state of Lower Saxony holds 20% of the voting rights, and under the Volkswagen Act the resolutions that would ordinarily need three-quarters of a shareholder vote need more than four-fifths here — so the state commands a blocking minority that not even Porsche SE's 53% can overrule. Its minister-president, Olaf Lies, who sits on the supervisory board, said afterwards that closing plants "is not a plan for the future." Whether that adds up to a veto over any particular closure is murkier than the shorthand suggests: the Act's two-thirds board requirement is written for building and relocating plants, and does not name closures at all.

Notice what is actually being argued about in Wolfsburg. Not whether the forecast was wrong. Everybody agrees it was wrong. The fight is over who pays for it — and it is being conducted by people who, in 2018, all read the same beautiful, consistent, twenty-year line, and all believed it, because there was nothing in it not to believe.

The honest doubt

Here is what this essay is not allowed to tell you: don't extrapolate. Because extrapolating is, on almost every day, the right thing to do — and the evidence for that is far stronger than the evidence for cleverness.

Philip Tetlock spent two decades collecting 82,361 forecasts from 284 experts, and found that they "usually lost to simple extrapolation algorithms." Dragging a ruler across the past beat the professionals. Worse for our purposes: the experts who did worst were the confident "hedgehogs" — the ones with one big idea, forever announcing that the trend was about to break. In the great forecasting competitions, the same lesson: across 3,003 time series, Makridakis and Hibon concluded that "statistically sophisticated and complex methods do not necessarily provide more accurate forecasts than simpler ones." In economics, Meese and Rogoff showed in 1983 that a naive random walk beat the structural exchange-rate models of the day. And the end of Moore's law has been announced, with excellent reasoning, in 2000, by Gordon Moore himself in 2005, and by DARPA's microsystems director in 2013 — while TSMC began mass-producing 2-nanometre chips in 2025.

So the people who spent twenty years saying Chinese growth would stop were wrong for twenty years. The turkey is right on almost every day of its life; that is what makes it a turkey and not an idiot. Anyone who bet against Volkswagen's China line in 2006 lost, and lost again in 2007, and kept losing until they had no capital and no credibility left to be right with.

The claim here is much narrower, and it survives all of that. It is not don't extrapolate. It is: do not let the length of the trend become the strength of your belief in it — because those two quantities are related backwards. The one thing your instrument cannot measure is the only thing that will hurt you, and the instrument grows most confident precisely as that thing approaches. Extrapolate. Just stop treating your error bars as though they had a vote on whether the world stays the same.

How to read a line that has only ever gone up

In January 2025, Volkswagen's China chief set a target: 3.5 million vehicles a year and about 15% of the market by 2030. It may well be the right plan. It is a forecast made by the same company, with the same tools, about the same country, one year after the line broke — and it will read, in either direction, as obvious in hindsight.

You can run the whole thing yourself: The Turkey. Grow a market for twenty years, watch the model's band close in around the future, watch a genuine backtest come back clean — then let the world change, and see the year you are standing in scored in the standard deviations of the forecast you wrote the night before. Turn one dial, how steady the good years were, and the identical collapse is a 3-sigma disappointment or a 24-sigma impossibility. Nothing about the world changed. Only how calm its past had been.

It has two siblings, and between them they cover the three ways a forecast goes wrong. The Monte Carlo Fan shakes the path around a growth rate you claim to know. The Fog Behind the Number shakes the rate itself. The Turkey leaves both alone and breaks the world instead — the uncertainty that no honest error bar has ever contained. And when the cutting is finished and someone explains why it was always obvious, Defensible If It Loses is waiting.