Flight Demand Atlas

I wanted a map that shows where people want to fly and where airlines actually put seats, so I could spot places that are underserved, starting with Puerto Rico, where tickets often feel expensive. Before drawing anything I needed to know what drives demand in the first place, so I wrote down seven hypotheses and one gate in advance: if the model could not beat a naive forecast on data it had never seen, the "gap" layer would not go on the map.

Everything here is public. The US Department of Transportation publishes a 10% sample of domestic airline tickets every quarter since 1993 (DB1B) and a monthly count of every scheduled flight segment (T-100). Both are US government works in the public domain. The BTS website blocked my build machine, so I pulled the same tables from two public GitHub mirrors and checked them against BTS press releases: 2019 scheduled domestic passengers came out at 818.7 million against 811.4 million reported, and December 2024 at 72.4 million against 72.1 million. County population and income come from the Census ACS 2015–19, Puerto Rico's municipio income from ACS 2013–17, and the country comparison from Airbus and the Maddison Project through Our World in Data.

An airport has no residents, so I built catchments. Each county's people are shared among nearby airport markets in proportion to seats divided by distance squared, a Huff model whose settings I fixed before looking at results. Airports less than 60 km apart, like Orlando and Sanford, count as one market. A trip belongs to the place where its ticket starts, which matters later: one-way tickets make leisure places look busier than they are.

For places, I regressed trips per resident on income, population, distance to a large hub, fares and visitor share, with robust errors and checks for collinearity. For routes, I fit a Poisson gravity model on 15,576 city pairs and tested it three ways. On pairs it never saw it explains 93% of the variation, every year from 2015 to 2024. As a next-year forecast it loses to "same as last year". And its "gap" did not predict where airlines launched routes in 2025 (AUC 0.48, a coin flip). That last test failed the gate.

When the first version was done, six independent reviewers, each playing a different specialist (a network planner and an econometrician among them), tried to break it without seeing my reasoning. I changed the page because of what they found.

I worked with Claude, Anthropic's AI model, as my analyst. It found the data mirrors, wrote the Python pipeline (pandas, statsmodels, scipy) and the D3 charts, ran the tests and ran the review as separate agents. I chose the questions, the scope (the US plus Puerto Rico), what counted as a pass or a fail, and what this page should say, and I checked the results against what I know from flying in and out of San Juan.

• The gap layer. My headline idea was to color places by actual ÷ predicted demand. It mostly restated whether a nonstop already existed, and it did not predict new routes. I withdrew it.
• The missing-routes list. 395 of the 617 "missing" pairs already had seasonal nonstops.
• Puerto Rico's income claim. Every PR market earns less than the poorest mainland one, so "PR flies twice what its income predicts" was an extrapolation. I dropped it.
• A circular calibration. The best-fitting catchment setting used the outcome itself, so I went back to the fixed setting.
• A stale fact. One version listed Cleveland–San Juan as a nonstop because it used 2024 schedules; Frontier paused it in January 2026. Every route's status is now checked against the latest T-100 month (May 2026) and airline news.

San Juan's international traffic (12.5% of SJU passengers in 2025) is not in BTS data, and two of my guesses about why Puerto Rico flies less, an older population and less business travel, are still untested. Those come next.