Auction Board

A phone-first draft companion for YUL to YYZ to SEA to SFO: a 14-team Yahoo head-to-head 9-category league with a $200 live auction (Mon 19 Oct 2026, 3 pm Pacific / 11 pm London).

The app is organised around getting ready for the auction:

• Prep: days to the draft, a four-step checklist with progress, and what the Yahoo room is likely to get wrong (players to let others overpay for, bargains to target).
• Board: your price for every player, with what Yahoo drafters usually pay next to it (green is a bargain, red is pricey). Tap a player to move his price with − / +, tag him Target or Avoid, see why he's priced that way, and adjust his games or minutes.
• Yahoo: the list to copy into Yahoo's pre-draft values: Yahoo's default crossed out, your value beside it, biggest changes first. Tick each one off; a price that changes after you entered it comes back to re-enter.
• News: changes waiting for your OK, each with what it does to the price. Applying an absence also gives teammates their measured share of his minutes.
• Draft: on the night, type the nominated player to see your bid limit (adjusted for the money left in the room), then record the sale.
• ⚙ Settings: team names, market prices, backup, and How the numbers work.

How good are the values?

Setup → Data & value quality shows both tests; re-run them with node scripts/backtest.mjs and node scripts/simulate-h2h.mjs <season> --projected.

• Projections (scripts/backtest.mjs → data/eval.json): each of 2022-23 to 2025-26 projected before it started, valued, and compared with what happened. Rank correlation: last season repeated 0.54, stats model 0.60, ESPN preseason 0.68, 75% ESPN + 25% stats model 0.68 with the smallest dollar miss ($6.81). That blend is what the app uses.
• Formula (scripts/simulate-h2h.mjs → data/eval-h2h-*.json): one team drafts from preseason projections with each formula against 13 plain-z-score teams, then the real season's weekly stats are replayed head to head. G-scores (z-scores that count each category's week-to-week noise, data/model/tau.json) won 66%, 66% and 61% of weeks in 2025-26, 2024-25 and 2023-24; capping z-scores did nothing. G-scores are the default; plain z-scores are a switch in Setup.
• Limits: snake draft as a stand-in for the auction, no position limits, everyone on a roster plays.

News pipeline (twice a day)

GitHub Actions, 1:30 pm and 11 pm Eastern (.github/workflows/data.yml)
  collect   ESPN projections, depth charts, injuries, headlines; FantasyPros consensus and player news
  Jev       reads every new text: absence length (tested 95% on hand labels) and 29 flags (96.5%)
  build     projections, ESPN bias corrections, depth moves and big projection moves
  signals   data/signals.json: what may need a judgement
Claude routine, 2:15 pm and 11:45 pm Eastern (jobs/routine.md, jobs/news.md)
  sizes the flagged items into proposals in data/adjustments.json, or skips them with a reason
You, in the app's News tab
  accept or reject; an accepted absence also gives its minutes to teammates

What the tests found (Setup → Data & value quality has the numbers):

• Jev's flags on preseason news don't beat ESPN. 2,274 archived Rotowire preseason notes (2022-23 to 2025-26, Wayback Machine) flagged by Jev, fitted against ESPN's projection misses one season held out at a time (scripts/fit-flags.mjs). Per-36 rates: nothing above noise. Games: the fit found effects pointing the wrong way (a minutes limit "adding" games), a healthy-player proxy, so a direction check rejects them. ESPN already prices preseason news in. Jev's job is therefore to read news fast and route it (absences, role changes) to proposals, not to move numbers itself.
• Kept: ESPN's average misses. ESPN projects ~1.3 minutes a game too many and assists per 36 slightly low, every season; corrected in the build.
• Teammates gain when a starter sits (scripts/fit-cascade.mjs, three seasons of weekly logs): each same-position teammate picks up ~10-15% of his minutes, others ~7%, with a small scoring-rate rise.
• Playoff weeks 18-20 (mid-Feb to 21 Mar): players 32+ play 4-8% fewer games than before, under-25s 3-5% more; a small availability factor.
• Uncertainty: 8 in 10 drafted players finish within about -$11/+$10 of their projected value; flag tiers didn't change that, so the app shows one range.

How values are made (lib/value.mjs)

1. Projections (scripts/build-players.mjs): the last two seasons' per-game stats from Basketball Reference, weighted by games played (the older season at half weight), with a small age nudge. Projected games = recent games regressed toward 70, capped at 78.
2. Category scores: a per-game z-score in each category, measured against the 182 players who will actually be drafted (14 teams × 13 roster spots). FG% and FT% count by impact (makes above an average shooter on the same attempts), so volume matters. Turnovers count against a player.
3. Availability: value per game above a replacement player × share of the season played. With a 2-adds-a-week cap you can't stream around injuries, so missed games cost real value.
4. Dollars: $2,800 in the room, minus $1 for each of the 182 spots, leaves $2,618. That is split in proportion to value above replacement, so the 182nd player is worth $1.
5. Live inflation: (money left − $1 × open spots) ÷ (value above $1 of the best players still available). If early overpays leave less money for the same talent, everyone left gets cheaper, and the reverse.

Data

.github/workflows/data.yml runs every six hours on GitHub: it fetches stats (daily), rebuilds projections, fetches ESPN injuries and headlines, runs the tests and commits data/. Start it once by hand from the Actions tab.

• data/positions.json: { "<bbref id>": ["PG","SG"] } to match Yahoo's position eligibility where it differs from Basketball Reference.
• data/extra-players.json: rookies and anyone without NBA stats, written as per-game projections.
• data/adjustments.json: proposals from the news job (append only).

Your ranks and draft are stored in the browser (localStorage). Use Setup → Backup to move them between devices.

Run locally

python3 -m http.server   # then open http://localhost:8000
node --test tests/*.test.mjs

It deploys as a static site: import the repo in Vercel, with no build command.