All Models Are Wrong
Settled science isn't supposed to be wrong. So why do I keep reaching for models I know are?
Neil deGrasse Tyson says settled science is never later found false. Caloric, the ether, and a thermostat say otherwise — and point at what we actually want from science.
Start here, then follow it as deep as you like.
Forecasting Citi Bike demand
A decision layer for one-day Citi Bike demand: conformalized quantile forecasts that restore 80% interval coverage, a 90th-percentile newsvendor rule for rebalancing, and an ops memo that turns uncertainty into a truck plan.
Which forecast should you trust?
A granularity-and-horizon audit of ten demand-model families on 110M Citi Bike trips: the daily champion is the worst annual accountant, nothing honest beats "same day last year" at 90 days, and equal information makes a zero-shot foundation model tie a tuned tree. Final champion: the arithmetic mean.
Citi Bike weather station explorer
A MapLibre station explorer for 2023 Citi Bike demand: filter by temperature band, rain, time of day, and day type to see which parts of the network move with the weather.
Who should get the ad?
A/B testing done properly on a real 14M-user randomized experiment: randomization audit, doubly-robust ATE (the naive estimate was 13–25% inflated), T-learner uplift models validated on held-out randomization, and a targeting policy keeping 77% of the lift at 30% of the cost.
Orbit Lab
An interactive applet that lays a perfect-circle model over a planet's true orbit and scores the fit live — R² and RMS update as you switch planets and toggle a heliocentric/geocentric origin.
Holding Orbits Up to the Light
Two thousand years of planetary models, and what modern data does to them.
Rediscovery
A botched classroom experiment that was supposed to prove conservation of mass — and accidentally rediscovered the buoyant force instead.