Skip to content

3 subjects · 6 studies

Work

The subject is variable; the method is not. Each subject below is here because it has published work, and each piece states what it is built from.

Football

All 4

Dashboard · refreshed weekly

The season in progress, over every season before it

Booking discipline in 11 leagues, one club at a time: its index match by match, with the interval, over the band of every completed season. Clubs are alphabetical and nothing is ranked. The league, club and season are in the address.

2026-27 · 194 clubs · 3,100 completed team-seasons behind it · through 2026-10-21

Football03L1 (descriptive)

The lever nobody pulls

Whether a foul is punished depends enormously on where and when it happens. A foul in your own defensive fifth is carded 31.4% of the time; the same offence in the attacking fifth, 8.9%. A foul in the opening quarter of an hour is carded 6.5% of the time; after ninety minutes, 25.6%.

Published
Figures
4 figures
Football04L3 (confirmed pattern)

How much of a booking index is real?

A club's yellows-per-foul rate over one season is mostly noise. 15% of the spread between clubs is a real club property; the other 85% is the Poisson accident of one season's bookings.

Published
Figures
2 figures
Football02L2 (hypothesis, adjusted)

A club that fouls with impunity, and the four ways I was wrong about it

Across eleven European leagues (58,013 completed matches, 115,984 team-matches, and 246 clubs with 150+ matches), raw yellow-cards-per-foul ratios make almost every dominant domestic club in Europe appear leniently officiated (r = −0.371). That apparent impunity is a compound artifact of three confounders: pre-match win expectation (1.098 heavy underdog vs. 0.856 heavy favourite median across all eleven leagues), non-foul conduct cards (dissent, delaying restarts, simulation) entering as a positive affine intercept (11.1%–38.8% league-mean share; 13%–47% across league-seasons), and fixture-level competitive intensity shared by both teams on the pitch (r = +0.283 on opponents' booking indices; 11 of 11 leagues rising across opponent-strength terciles 0.938 to 0.993 to 1.088).

Published
Figures
8 figures
Football01L1 (descriptive)

What free football data can still tell you

Public European football repositories exhibit severe longitudinal truncation, unrecorded rule differences in second-yellow accounting, non-foul conduct cards in the yellow-card numerator, and missing referee identifiers across 9 of 11 European leagues. Across 286 completed league-season files in 11 top divisions (2000-01 to 2025-26), only 179 files (62.6%) carry foul and card statistics (58,013 completed matches), while the remaining 107 files return HTTP 200 and parse cleanly into tables containing zero match events. Referee identity is recorded in only 55 of 286 files — continuously in England and Scotland, for two early seasons in Germany and Italy, and never in the other seven leagues.

Published
Figures
3 figures

Live and unresolved: The Portuguese season as it stands, updated .

Energy

Open the dashboard

Dashboard · refreshed daily

A day on the grid, in 5 countries

Pick a country and any day since 2015 and see the price through the day, generation by fuel and demand; cross-border flow too for Portugal and Spain. The day is in the address, so a post can point at one afternoon.

4,268 days · 102,418 hours · through 2026-09-07

AI in operations

All 2
AI in operations02L2 (hypothesis, pre-registered analysis)

Four frontier models on the same support queue: what separates them, and what does not

A year of these rounds has taught me that most model comparisons are run on the work a system already does well. Ours was too. On the tickets our Support team closed without help, four frontier models are interchangeable: they name the root cause between 87% and 96% of the time and you could pick on price alone.

Published
Figures
7 figures
AI in operations01L2 (hypothesis, pre-registered analysis)

Putting an agent on the incident queue: what I measured before I trusted it

I run Applied AI and Support at Yugabyte. Both halves of that title report to the same outcome: how well, and how soon, we help a customer who needs us. So when we put an AI agent in front of the support engineers, the question was never "which model scores highest". It was: can an engineer act on what this agent says, how often will it cost them time, and can we tell in advance.

Published
Figures
10 figures