How this was built
A bracket, with its uncertainty intact.
The planning choices and evidence behind this independent tournament explorer.
The request
Recreate FiveThirtyEight’s 2015 dashboard, then add a selector for every year from 2016 onward. The user later requested another documented probability model for the years without FiveThirtyEight forecasts.
This is a real request and development account, not a simulated student interview. The supplied reference selected men’s basketball; a subsequent request added a Men/Women toggle and women’s tournaments for the same years.
Sources and reconstructed updates
The expanded request adds complete opening fields from Ken Pomeroy, Bart Torvik, FiveThirtyEight, ESPN BPI, numberFire and Silver Bulletin, where recovered. Silver is limited to women’s 2024. The selector covers both competitions from 2016 through 2026; 2020 remains cancelled.
There are 37 complete source fields. A bracket-specific matchup model fits each source’s opening advancement probabilities, then recomputes the remaining odds after each day’s completed games. FiveThirtyEight’s source selection instead shows its published historical snapshots. Other sources and both blends use app reconstructions, not archived publisher updates. Missing First Four odds and rounded values require disclosed assumptions. Original archives remain in the downloadable data.
An equal-source average requires at least two sources. The logistic ensemble requires prior-year observations for every current input and never trains on the selected year’s outcomes. Its comparison accordion uses chronological held-out seasons and reports both improvements and worse performance.
The recipe
Browser App Builder provides the planning, build, evaluation, and deployment preparation. Campus Designer supplies local fonts and a consistent palette. Native HTML and SVG handle the bracket and sortable table; Apache ECharts draws the selected team’s forecast history, with a readable table equivalent.
The app is a static Vite build with pinned dependencies. Visitor data, fonts, and scripts are served locally, without runtime APIs or telemetry.
What was checked
The source expansion adds independently calculated bracket cases, future-outcome isolation, chronological fitting checks, and an audit of all 60 source/blend options and 757 daily snapshots. Rounding adjustments are recorded per source. These checks establish internal consistency, not a guarantee of predictive accuracy.
The original men’s release passed 80 of 80 automated data and model checks on the configured Mac. Checks include original source probabilities, tournament topology, all seven round totals, First Four sharing, independently calculated model cases, and Oregon’s 2021 advancement over VCU without a played game.
An independent source/model review identified six issues that were corrected: stale finalist labels, inconsistent accessible descriptions, VCU’s no-contest status, replay reset, First Four loser overlays, and chart restoration after browser Back. Browser review also prompted corrections to regional navigation and tiny-probability announcements. Numerical invariants and source review do not establish predictive accuracy, complete accessibility, or novice usability.
Follow the sources
- FiveThirtyEight’s data repository
- T-Rank’s rating and log5 explanation
- T-Rank’s archive guidance
- NCAA championship history
- Source-specific data credits and licenses
- Library and font notices
Source, planning, tests, and evaluation
The explorer’s source disclosure links the exact data used for each year. Full planning and evaluation records stay in the project’s source handoff, outside the website.