Your dashboard is a model with better fonts. Every tile on it, from revenue by channel to customer satisfaction, rests on assumptions about what counts, what causes what and what will stay the same. Ask about them in a meeting and someone will shrug and quote George Box: “all models are wrong, but some are useful.”
He usually means: stop asking awkward questions about my number. That is close to the opposite of what Box meant. His sentence is a demand to know which assumptions you are relying on and which of them could hurt you, and nobody prints those in 14-point type next to the green arrow.
Who said “all models are wrong, but some are useful”?
George E. P. Box was born in Gravesend, England, in 1919, worked as a statistician at Imperial Chemical Industries, and in 1960 went to the University of Wisconsin to create its statistics department. He died in Madison in 2013.
The first half appears in “Science and Statistics,” his 1976 paper in the Journal of the American Statistical Association. It is the written version of the R. A. Fisher Memorial Lecture he gave in St. Louis in 1974, about Fisher, who was also his father-in-law. Two short paragraphs on page 792 say:
Since all models are wrong the scientist cannot obtain a ‘correct’ one by excessive elaboration.
Since all models are wrong the scientist must be alert to what is importantly wrong. It is inappropriate to be concerned about mice when there are tigers abroad.
The full sentence arrived in 1979. In “Robustness in the Strategy of Scientific Model Building”, a chapter of the book Robustness in Statistics, Box used “All models are wrong but some are useful” as a section heading on page 202. His example was the ideal gas law, PV = RT, which is not exactly true of any real gas and is still enormously useful. The only question of interest, he wrote, is “Is the model illuminating and useful?”
In 1987 he and Norman Draper gave the most-quoted version in Empirical Model-Building and Response Surfaces: “Essentially, all models are wrong, but some are useful” (page 424). The context was a polynomial approximation, which is no less useful for being one.
Quote sites also credit that book with a longer line about how wrong a model must be before it stops being useful. I couldn’t find it in a primary source, so it stays out. A post about trusting models is a poor place to trust a quote site.
What did George Box mean?
Read the 1979 chapter from the start and the meaning is hard to miss. In its first paragraph Box observes that even a procedure people use without stating any model “implies some unstated model,” and that there is “great virtue in bringing into the open” the assumptions behind it. A rule of thumb, a forecast, a weighting scheme: each has a model inside it, whether or not anybody wrote it down.
In the 1976 paper he named the disease that follows when nobody does. He called it “cookbookery”: forcing every problem into one or two routine techniques, with too little thought for the real objectives “or to the relevance of the assumptions implied by the imposed methods.” I can’t think of a better description of most corporate reporting.
So the sentence has two halves, and both are demands. “All models are wrong” means you must know how yours is wrong. “Some are useful” means you must be able to say useful for what, and how you would notice when it stopped being so. Box was asking people to worry selectively, and the quote now gets used as permission not to worry at all.
Why is a dashboard a model?
A dashboard looks like a window onto the business. It is really a theory of the business.
“Marketing-sourced revenue” assumes you know which touch caused a sale. “Active users” assumes a definition of active that someone chose on a Tuesday and nobody has revisited since. A satisfaction score assumes that customers who answer the survey resemble those who don’t. A forecast tile assumes next quarter will behave like the last twelve.
None of that is wrong in itself, and Box would say it couldn’t be anything else. The trouble is that the assumptions are invisible, the fonts are confident and the meeting has 15 minutes. I made a related complaint about research in You killed market research: dashboards are good at showing that a number moved and silent about why. Here are four models whose assumptions failed in public, each with warning signs for anyone who asked.
Did a formula kill Wall Street?
In 2000 David X. Li, a quantitative analyst, published “On Default Correlation: A Copula Function Approach” in the Journal of Fixed Income. He borrowed the idea from actuaries, who had noticed that people tend to die sooner after losing a spouse. “Default is like the death of a company,” he later told The Wall Street Journal, “so we should model this the same way we model human life.” His Gaussian copula squeezed the chance that many borrowers default together into a correlation figure, and banks used it to price slices of collateralized debt obligations.
That correlation was the crucial assumption. Felix Salmon’s 2009 Wired cover story, “Recipe for Disaster: The Formula That Killed Wall Street”, pointed out that it was estimated from credit default swap prices, which had existed for less than a decade, a period in which house prices soared. A model calibrated on a boom assumed the boom’s relationships would hold. When home prices fell across the country at once, the correlations jumped.
The fair hearing belongs to Li, because he gave it himself. In the 2005 Wall Street Journal article, he said “it’s not the perfect model” and worried that “very few people understand the essence of the model.” Then came the line every model owner should frame: “The most dangerous part is when people believe everything coming out of it.”
The same article described a warning shot. When Standard & Poor’s cut General Motors and Ford to junk in May 2005, hedges built on the model’s simplest assumption failed and some hedge funds lost heavily. The model let traders adjust that assumption; some had simply trusted it.
In August 2007 Goldman Sachs’s chief financial officer, David Viniar, explained his firm’s hedge fund losses to the Financial Times: “We were seeing things that were 25-standard deviation moves, several days in a row” (quoted in How Unlucky is 25-Sigma?). Under a normal distribution, a single 25-sigma day shouldn’t happen once in the lifetime of the universe. When a model calls something impossible and it happens several days running, it is the model that has failed.
Google Flu Trends: part flu detector, part winter detector
In February 2009 Google and CDC scientists reported in Nature that search queries could track influenza-like illness with a lag of about a day, well ahead of official surveillance. It became the poster child of big data.
The assumption buried inside was that the relative volume of certain searches stays tied to how many people are ill. A 2014 analysis in Science by David Lazer and colleagues showed how that broke.
The original model searched 50 million candidate terms for the best fit to 1,152 data points, an open invitation to find terms that rise every winter for reasons unrelated to flu. Its developers had to weed out searches about high school basketball. It then missed the nonseasonal 2009 H1N1 pandemic. In the authors’ phrase, it was “part flu detector, part winter detector.”
The revised model failed in a subtler way. Starting in August 2011 it overestimated flu prevalence in 100 of 108 weeks, and in February 2013 it was predicting more than double the share of doctor visits for influenza-like illness that the Centers for Disease Control and Prevention reported. Even three-week-old CDC data did a better job of estimating current flu levels.
A likely culprit was Google itself: it kept changing its search engine, suggesting related searches and offering diagnoses for symptoms, and so changed the very data its model read. In August 2015 Google shut down its Flu Trends website.
Lazer and his coauthors were fair about it. Combined with CDC data and recalibrated, the signal was useful. Used alone, as a substitute for the slow data it was built to predict, it wasn’t.
That lesson applies to every dashboard that has quietly swapped a measure for a proxy. Goodhart’s law is the darker version, in which people paid on the proxy move it on purpose.
Why do polls miss? The assumption about who answers
A poll is a model too. Its central assumption is that, after weighting, the people who answered resemble the people who didn’t. The AAPOR review of 2016 found the national polls basically right: they had Clinton about 3 points ahead, and she won the popular vote by 2. State polls in the Upper Midwest underestimated Trump, partly because voters decided late and partly because many polls didn’t weight by education, although college graduates are more likely to answer surveys and in 2016 were more likely to vote for Clinton.
The industry fixed the assumption it had found. In the final two weeks of 2020, 92% of the state presidential polls that disclosed their adjustments accounted for education. According to AAPOR’s 2020 task force, the result was the largest polling error in 40 years for the national popular vote and in at least 20 years for state-level estimates. Polls were too favorable to Biden by 3.9 points nationally and 4.3 points in the states.
The task force ruled out late deciders and education weighting and said a conclusive explanation was impossible with the data available. Its leading suspect was the assumption no poll can check from inside: that the people who decline to take part are like the people who agree.
Did the ad really cause the sale?
Marketing dashboards bury a different assumption: that the ad someone clicked before buying caused the purchase. eBay tested it.
In experiments published in Econometrica in 2015, Thomas Blake, Chris Nosko and Steven Tadelis found that the returns from paid search were a fraction of non-experimental estimates. Ads on brand keywords, queries containing “eBay,” had no measurable short-term benefit: people searching for eBay found eBay anyway. For other keywords, most of the money went on frequent users who would have bought regardless, so average returns were negative.
A larger test came from Facebook. In a 2019 Marketing Science paper, Brett Gordon and colleagues compared 15 large randomized advertising experiments with the observational methods the industry commonly uses. The observational methods often failed to recover the experimental effects, even with rich data on users’ demographics and behavior. In my experience, most attribution dashboards rest on the same kind of methods.
What does “useful” mean in practice?
Box never said a model had to be right. He said you had to know how it was wrong. In practice I think that comes down to five questions:
- What does it assume? Write the assumptions down in plain words. If nobody can, you are running what Box called an unstated model, and trusting it blind.
- Does it beat a dumb benchmark? Google Flu Trends lost to three-week-old official data. A forecast that can’t beat “same as last year” is decoration.
- Has it met an experiment? Search ads looked profitable until someone switched them off. Where you can randomize, do; where you can’t, at least hold something back.
- Who can change the data underneath it? Google changed search, customers change habits, and staff change whatever they are paid on.
- Which of its errors would hurt? That is the mice and the tigers. A copula slightly off on a quiet day is a mouse. A correlation that jumps when house prices fall is a tiger.
None of this needs a statistics degree. It is the ordinary critical thinking question, “How do I know that?”, asked of a number instead of a person. The same goes for AI respondents that pose as real people.
My take
I like models. I have built a fair number, and a few of them were useful. What I dislike is how Box’s sentence gets used in meetings, as a shrug that ends the conversation. He meant it as the start of one.
Organizations make this worse in a predictable way. The people who built the model move on, the dashboard stays, and the new crew inherits the numbers without the assumptions. In Submariners in space I described task forces that report activity instead of change. A dashboard nobody can explain is the same problem with a login screen.
My rule is simple. Every number that moves money should come with a footnote, in words, saying what has to be true for it to be right. If nobody can write that footnote, the number may not be wrong yet. It just hasn’t been caught.
Questions people ask
Who said ‘all models are wrong, but some are useful’?
The British statistician George Box. He wrote ‘all models are wrong’ in a 1976 paper in the Journal of the American Statistical Association, used the full sentence as a section heading in 1979, and repeated it in his 1987 book with Norman Draper.
What does ‘all models are wrong, but some are useful’ mean?
Every model simplifies reality, so the real question is whether it is wrong in ways that matter for your decision. Box wanted modelers to know their assumptions and to worry most about the ones that could do serious damage.
Is a dashboard a model?
Yes. Every number on it rests on choices about definitions, data sources, attribution and what to leave out. Those choices are assumptions, and they can fail the way a statistical model’s assumptions do.
How do you know if a model is useful?
Write down its assumptions, compare it with a simple benchmark, test it against an experiment or fresh data, and ask who can change the data it depends on. A model that can’t beat last year’s number isn’t earning its keep.