I make data make sense.

1. 1

   Goals are raining down.

   About 3 goals every game. No World Cup has scored like this since 1958.

2. 2

   Every shot is a chance, and chances can be counted.

   A tap-in almost always scores. A shot from far away almost never does. Add up what every shot is worth and you get the goals that SHOULD happen.

3. 3

   The chances say 90. The scoreboard says 109.

   Strikers scored 19 more goals than their chances were worth. That is the biggest gap ever measured at a World Cup.

4. 4

   So is it magic? Probably not.

   Luck like this usually fades. When someone shows you a hot streak, ask what the chances were worth.

Sources: Total Football Analysis, 3 July 2026 (109 goals vs 90 expected, about 3 goals per game, most since 1958). Historic rates: FIFA. xG (expected goals) = how often that exact shot scores.

1. 1

   The total looked bad.

   Of every 100 men who applied, about 44 got in. Of every 100 women, about 35. Newspapers were ready.

2. 2

   Someone split it by department.

   Look inside each department and the story flips: women got in at the same rate or better in most of them.

3. 3

   Women had picked the harder doors.

   More men applied to departments that accept lots of people. More women applied to the toughest ones. The total mixed those choices into a fake story.

4. 4

   Averages can lie. Split them first.

   One number for everything can point the wrong way. A data scientist always asks: what happens inside each group?

Sources: Bickel, Hammel and O’Connell, Science (1975): 12,763 applicants to UC Berkeley, autumn 1973. Overall rates 44% of 8,442 men, 35% of 4,321 women; department rates rounded.

1. 1

   One number told everyone to cut spending.

   The claim: when a country’s debt passes 90% of its economy, growth turns negative. Politicians everywhere quoted it.

2. 2

   A student tried to rebuild it. He could not.

   For a class project, Thomas Herndon re-ran the numbers again and again. His answer never matched the famous one. So he asked for the original spreadsheet.

3. 3

   The formula skipped five countries.

   A formula had been dragged down the sheet, but not far enough. Five countries at the top of the list were simply never counted.

4. 4

   The honest number was positive.

   Fix the formula and the scary negative becomes plain positive growth. Same data, working pipeline. Check every step, especially the boring ones.

Sources: Reinhart and Rogoff, Growth in a Time of Debt (2010); Herndon, Ash and Pollin, University of Massachusetts Amherst (2013): the corrected average is +2.2%, not -0.1%; five countries excluded by the spreadsheet error.

1. 1

   Google tried to predict the flu.

   The idea: when people feel ill, they search. Count searches like “flu symptoms” and you can guess how much flu there is right now, weeks before the doctors report it.

2. 2

   At first, it worked.

   The model tracked the real flu numbers closely, season after season. It became the poster child for big data.

3. 3

   Then it missed. Badly.

   For 100 weeks out of 108 it guessed too high. At the 2013 peak it predicted more than double the real amount of flu.

4. 4

   Test it on data it has never seen.

   Any model can look brilliant on the past it learned from. Trust begins when it keeps working on new, unseen data. That is the whole audit.

Sources: Lazer, Kennedy, King and Vespignani, The Parable of Google Flu, Science (2014): the model overestimated in 100 of 108 weeks from August 2011, and predicted more than double the real level at the February 2013 peak (real counts: the US CDC).

1. 1

   Nobody knows exactly where a storm will go.

   Not even the best scientists with the best computers. The wind has not decided yet. Pretending otherwise would be a lie.

2. 2

   So they draw every path it might take.

   The computers run the storm forward many times, each run slightly different. Together the paths make a fan.

3. 3

   The fan becomes a cone that widens.

   Tomorrow they are quite sure. Five days out, much less. The cone grows about 8 times wider from day 1 to day 5, because honesty grows with distance.

4. 4

   And they tell you how often it works.

   The storm stays inside the cone about 2 times in 3. The forecasters print that on the forecast itself. That is what a trustworthy prediction looks like.

Sources: NOAA National Hurricane Center: the cone is drawn so the storm’s centre has stayed inside it about 2 of 3 times over the previous five years; the cone circles grow roughly 8 times wider from day 1 to day 5.

1. 1

   One page, one button, millions of visitors.

   The splash page asked people to sign up. Every visitor saw a picture and a button. Which picture? Which button? Opinions everywhere.

2. 2

   Instead of arguing, they split the crowd.

   Visitors were shown different versions at random: 6 pictures and videos, 4 button texts, 24 combinations. Random is the magic word: it makes the groups fair.

3. 3

   The boring option won. By a lot.

   Everyone expected a video to win. The winner was a plain family photo with a button saying “Learn more”. Sign-ups jumped from about 8 in 100 to about 11.6 in 100.

4. 4

   That one test was worth about $60 million.

   The better page meant millions more sign-ups, more volunteers, more donations. Nobody predicted it. The test found it. Test, do not guess.

Sources: Dan Siroker (the campaign’s analytics lead, later Optimizely), How Obama Raised $60 Million by Running a Simple Experiment (2010): sign-up rate 8.26% to 11.6% (+40.6%), about 2.9 million extra sign-ups, about $60 million in added donations, 310,382 visitors in the test.