Skip to main content
Read-only snapshot: model data as of 2026-07-16; paper desk record through 2026-10-22. No live data; paper units only.
CourtVision
Ask Scout

About CourtVision

A calibrated sports forecaster that publishes its own scorecard against the market, including losses, nulls and retractions. Its headline finding: the pregame model does not price games better than the market. It trails the Devigged close the closing line with the vig removed, i.e. the market's fair probability at the final pre-game price; the comparison forecaster for every pregame-vs-close row (see Devig / Shin devig and Closing line). on MLB moneyline and soccer over/under 2.5; on NBA moneyline the interval includes zero, so we cannot tell. Each headline figure names its source artifact and as-of date, and the evidence packet is the truth source.

The aim is a forecaster whose every number someone else can check against the market.

What it is

CourtVision makes probability forecasts for NBA, MLB, soccer and tennis games, before and during play. It scores them out of sample with the Brier score mean squared error of probabilistic forecasts (lower = sharper).: against the devigged closing price before a game, and against the live market during it. The goal is calibration: a 60% forecast should come true about 60% of the time. Every published number carries an as-of date, a source file and a verdict word taken from the evidence packet.

How it is built

One person directed coding agents to build it. The owner set the goals, the validation rules and the acceptance bars; agents wrote most of the code, tests and documentation. One owner decision, for example: the former headline return figure was withdrawn once its grader was found never to read the model. Counted from git log at commit 4f033b5a5 on 2026-10-21, 1,620 of the 6,556 commits on the public master branch (24.7%) carry a Co-Authored-By trailer naming Claude or Codex, and the true agent share is higher because older and automated commits carry no trailer (CONTRIBUTING.md). Results that failed later checks were withdrawn in public and are listed on the retraction record.

git log (public master)as of 2026-10-21descriptive_onlygit rev-list --count 4f033b5a5; git log 4f033b5a5 -i --grep='^Co-Authored-By:.*\(Claude\|Codex\)' --format=%H | wc -l

What this is not

  • It is not a betting product and has no order path. Every desk figure is in paper units.
  • It does not price games better than the market. Pregame, the model trails the devigged closing price, and on NBA moneyline the interval includes zero, so we cannot tell.
  • The forward paper series has no calibration verdict yet. The receipts say why.
  • Pages show a dated static snapshot, not live data.

What is next

The forward paper series keeps running until each cell reaches its 30-game floor; the NBA fitted arms need a new dated seal. Computer-vision tracking stays paused.

Methods

Author and contact

Open to data-science / forecasting, quant-research, ML-engineering and analytics-engineering roles.

Built by Neel Shah, who has three years of data-science coursework at the University of Iowa.