NFL tight end touchdown forecasts

Estimates each NFL tight end's chance to score a touchdown from usage data, compares it with sportsbook prices, and grades the picks after the games.

Site: https://mattrudy13.github.io/te-tracker/

At a glance

• What it does: reads every game's box score and play-by-play, measures each tight end's usage (targets, target share, red-zone targets by distance), and turns that into an anytime and first-TD chance. Picks in six categories are recorded from live odds before kickoff and graded afterward.
• How it's checked: the model was calibrated against de-vigged market prices (the error fell from 11.2 to 6.4 percentage points on the first snapshot). Rules that leave players out are tracked in separate "shadow" records, so what each rule costs is measured rather than assumed.
• Live record (as of Oct 21, 2026): 11–39 on 50 official picks over weeks 5–6, −11.0 units (−22% ROI). Most picks are long odds, so two weeks says little; the rules are reviewed after week 9.
• Built with: JavaScript (Node), GitHub Actions on a schedule, GitHub Pages, ESPN's public API and The Odds API.

Details

• Best Bets (index.html, the home page): this week's TD picks in six categories: most likely to score, best value, longshots, due for a TD, fades, and first-TD value. Each pick shows the best price, model %, market %, edge, expected profit per $10 bet and a one-line reason. Below that is a track record of every past pick, graded after the games (W–L–void, hit rate, units, ROI, and closing-line value). Players listed as Out, Doubtful or on IR are never picked.

• Matchups (matchups.html): this week's games (next week's once Monday night is final) with kickoff, network, spread, total and implied team points. Expand a game to see both teams' tight ends:

  • Usage: targets per game and target share
  • Receptions: per game, next to the line and the best over/under prices
  • Touchdowns: TDs, red-zone targets, and the model TD % against the best book price, with the edge
  • how the opposing defense has fared against TEs

  Finished games show the TE box score with the pregame TD price and receptions line, and how each one landed. ?open=<gameId>,… keeps games expanded.

• Leaders (leaders.html): usage and touchdown leaderboard, per game or season totals: targets, target share, receptions, catch %, yards, yards per target, TDs, red-zone targets, xTD and TD − xTD, plus this week's receptions line and best anytime-TD price. Filter by team, week range, minimum targets or your ★ watchlist.

• Touchdowns (touchdowns.html): this week's anytime-TD or first-TD odds for every TE (an Anytime / 1st TD toggle, ?mkt=first; sorted by model-vs-book edge), red-zone usage (targets inside the 20 / 10 / 5), expected TDs (xTD) vs actual, and the TE touchdown rate by field position that the model uses.

• Player (player.html?id=<espn id>, click any player name): this week's game with the receptions line (best over/under, how often the player has gone over it), anytime-TD and first-TD model vs book; season usage; touchdown profile; charts of targets, receptions against the line, red-zone targets and yards; and a game log.

Filters live in the URL, so any view can be shared. The watchlist is saved in your browser.

How it works

No server and no build step: GitHub Actions write JSON into data/, and GitHub Pages serves the HTML plus that JSON.

update-data.yml    every 3 h Thu–Mon + Tue   build-data.mjs (ESPN) → grade-picks.mjs
snapshot-odds.yml  Thu 11 AM + Sun 4 AM ET   build-data.mjs → fetch-odds.mjs (The Odds API) → record-picks.mjs
check-assets.yml   on push                   stamp-assets.mjs --check
                                    ↓
          data/2026.json · odds-2026.json · picks-2026.json → GitHub Pages

File	What it is
index.html, matchups.html, leaders.html, touchdowns.html, player.html	The pages. Each keeps its page logic in an inline script
model.js	Data and model code (aggregation, TD model, injuries, best bets), shared by the pages and the scripts
common.js	Browser helpers: data loading, formatting, tables, URL state, watchlist
style.css	All styles (light/dark)
scripts/	build-data, fetch-odds, record-picks, grade-picks, stamp-assets (Node 20+, no dependencies)
data/	Generated JSON, committed by the workflows

scripts/build-data.mjs:

1. Reads the season calendar and every completed regular-season game.
2. Finds tight ends from the 32 team rosters, plus ESPN athlete lookups for players no longer on a roster (cached in data/positions-<season>.json).
3. Takes each game's box score (receiving, rushing, fumbles, and team totals for share stats) and parses play-by-play for red-zone targets and TDs. The box score only lists players with a stat, so each team's game roster (ESPN's core API) adds an all-zero row for every TE who played without one (about 40 a week, mostly blockers). Without them, part-time TEs' per-game rates ran high.
4. Saves the current week's schedule (state, score, network, lines) for the Matchups page, and the unplayed games' implied team points for the TD odds.
5. Saves each TE's injury status from the rosters (Questionable / Doubtful / Out / IR).

Games already in the file are reused, so a rerun only fetches new games. A game that can't be read (an ESPN error, or a game roster that isn't posted yet) is skipped and tried again on the next run. Picks on it aren't graded until it goes through.

.github/workflows/update-data.yml runs every 3 hours Thursday–Monday plus Tuesday morning (UTC) from September through February. It commits only when the data changed. You can also run it by hand from the Actions tab ("Run workflow"). Both data workflows build from the latest main, so re-running an old failed run works, though the next scheduled run catches up on its own.

Sportsbook odds

scripts/fetch-odds.mjs takes a snapshot of tight end props from The Odds API: anytime-TD and first-TD-scorer prices and the receptions over/under, from US-licensed books only (DraftKings, FanDuel, BetMGM, Caesars, BetRivers, ESPN BET, Fanatics…). Offshore books (BetOnline.ag, Bovada, BetUS, MyBookie, LowVig) are dropped before the best price and market median are chosen, so every price on the site can be bet at a licensed book. It writes data/odds-<season>.json. .github/workflows/snapshot-odds.yml runs it Thursday at 11 AM ET (before TNF) and Sunday at 4 AM ET. Both are early on purpose: GitHub often starts scheduled runs hours late. Games that have already kicked off keep their earlier snapshot. GitHub can start scheduled runs hours late (the first Thursday run started at 9 PM ET, after TNF kickoff), so TNF picks may come from an earlier snapshot. Each snapshot costs 1 credit per market per game (3 markets, ~15 games), so about 30–45 credits, ~85 a week and ~375 a month. That leaves ~125 for manual runs on the free tier (500/month). Each run logs the credits left.

The key is the ODDS_API_KEY repository secret, so it's only used inside the workflow. Set or replace it with gh secret set ODDS_API_KEY -R mattrudy13/te-tracker. To run locally: ODDS_API_KEY=... node scripts/fetch-odds.mjs. Without the file, the odds columns just show "–".

First-TD markets carry a 25–40% margin, so the snapshot also stores a de-vigged fairProb: each book's prices are divided by that book's total implied probability over every first-TD outcome in the game (books listing fewer than 20 outcomes are skipped, since a partial list understates the margin).

Sportsbooks post most TE props midweek, so a Monday-night or Tuesday run prices only a few players. Thursday's snapshot is the one that counts.

Where the odds show up:

• the Touchdowns table (Book, Market, Edge, Rec line; the 1st TD view swaps in first-TD prices)
• the Matchups page (Line, O / U, Model, Book, Edge before kickoff; pregame price and line vs result after; first-TD model and price as a second line under Model and Book)
• the Leaders table (Rec line, Book TD)
• the "This week" card on each player page

Best bets and the track record

The picks logic is bestBets() in model.js. That file is shared by the site and the Node scripts, so the Action computes exactly what the page shows.

Category	Rule (top 5 each)
Most likely to score	Highest model TD %
Best value	Expected profit > 0 at the best price, price under +400, model ≥ 15%, ≥ 2 books
Longshots	Price +400 or longer, expected profit > 0, at least 1 red-zone target, at least 2 targets in the team's last 2 games
Due for a TD	TD − xTD ≤ −0.8 with ≥ 3 red-zone targets
Fades	Price +250 or shorter and model at least 8 points below the market
First TD value	First-TD bet: expected profit > 0 at the best price, model ≥ 4%, ≥ 2 books, ≥ 1 red-zone target, ≥ 2 targets in the team's last 2 games

• scripts/record-picks.mjs runs after each odds snapshot, following a fresh stats/injury build. It saves the picks to data/picks-<season>.json. Picks lock at kickoff: Sunday's snapshot only replaces picks for games that haven't started. A replaced pick keeps the price it was first picked at (openPrice, openMarketProb), so the last snapshot before kickoff serves as the closing line.
• CLV (closing-line value) in the track record is the average move in the market's chance toward the pick (away from it for fades), in percentage points, from the first snapshot that had it to the last one before kickoff. Beating the close is a faster, less noisy sign of an edge than W–L. Picks seen at only one snapshot aren't counted, and picks dropped between snapshots aren't kept.
• scripts/grade-picks.mjs runs after every stats build. It grades final games: a TD (receiving or rushing) wins, a box-score line without a TD loses, and no game row (inactive, or dressed but never took a snap) is void. Profit is stored per $100 at the recorded price; the page shows it per $10 bet. Fades are graded right/wrong only. First-TD picks win only if the player scored the game's first TD (firstTd on the game row) and lose otherwise, including games with no TD. Reruns don't change graded picks.
• Until a snapshot records the week's picks, the page shows a live preview.
• When fewer than five players meet a category's rule, the list is filled with close calls: the next-best players from a looser pool, in lighter text with a "CC" label. They're recorded and graded as a separate shadow record that never counts toward the official one, to show whether the rules are too strict or too loose. A fade close call never repeats a real "due" pick, and vice versa.
• Fades aren't bets, so they're kept out of every Total / Unique bets row. Each fade shows ✓ No TD or ✗ Scored, and a line under each table compares how many faded players scored with how many the market and the model expected (a fade "wins" ~65% of the time with no skill, so wins and losses say little).
• Players the Longshots / First TD value usage rule (2+ targets in the team's last 2 games) cuts from the top five are recorded and graded in a third shadow table, Cut by the usage rule, to show whether the rule is costing winners. They never appear in the pick lists.

The touchdown model

• xTD credits each target with the league-wide TE TD rate for where the ball was snapped: inside the 5, 6–10, 11–20, or outside the red zone. Rates are this season's results blended with long-run priors, so the early weeks aren't driven by a few plays.
• Anytime TD %: expected TDs per game, shrunk toward a volume-only baseline (targets per game × a long-run 5.5% TE TD-per-target rate, weighted like 8 games of data), blended 70/30 with the last 3 games, plus a small rushing-TD term, with a floor of 0.02 expected TDs per game (so a TE who plays is never at 0%). That's scaled by the team's implied points against the week's average and run through a Poisson: P = 1 − e^(−λ). It's shown as a probability and as fair American odds.
• Calibration (Oct 5, 2026, first odds snapshot): with books de-vigged by ~12%, the error against the market fell from 11.2 to 6.4 percentage points (RMSE over 30 TEs) and the model's average matched the market's. Rechecked Oct 13 once zero-target games counted: the baseline went from 5% to 5.5% and the floor was added (from results: TEs with no prior targets scored in 2 of 100 games). Week 5 Sunday snapshot: RMSE 5.0 points, model average 0.8 points under the market. Constants live at the top of the TD model section in model.js.
• First TD %: the game's expected TDs are Λ = both teams' implied points × TD_PER_POINT (0.105, about 2.4 TDs on 23 points). If TDs arrive independently, the first one is the player's with chance λ/Λ, and there is one at all with chance 1 − e^(−Λ): P = λ/Λ · (1 − e^(−Λ)). TD_PER_POINT isn't calibrated against the first-TD market yet. Over weeks 1–4 a TE scored first in 10 of 64 games (16%); the model's TE share is about 20%.
• First-TD scorer: build-data.mjs reads ESPN's scoringPlays and flags the TE row that scored the game's first TD (firstTd: true). Games processed before this existed were filled in with node scripts/build-data.mjs --backfill-first-td.
• Injuries: players listed as Out, Doubtful, on IR or suspended get no TD % and are never picked.

Cache-busting

GitHub Pages lets browsers cache files for 10 minutes. The pages load model.js, common.js and style.css with a content-hash stamp (model.js?v=1a2b3c4d), so a new page never runs against a stale cached script. After editing any of those three files, run node scripts/stamp-assets.mjs before committing. The "Check asset stamps" workflow fails on a push if you forget.

Running locally

node scripts/build-data.mjs          # current season; or pass a year
python3 -m http.server 8000          # then open http://localhost:8000

Known limits

• ESPN doesn't publish snap counts or routes run, so usage means targets and target share.
• Red-zone splits come from play-by-play text and match box-score targets in ~95% of games (sometimes off by one).
• Injury status comes from ESPN rosters (refreshed every 3 hours and before each odds snapshot). "Missed last game" appears only for players without an injury tag, and means the player didn't play (or never took a snap) in the team's latest game.
• The TD model is calibrated against one snapshot. The Best Bets track record (first graded week: week 5) is the first real test of it.
• A defense-vs-TE matchup factor was tested and left out of the model: four weeks of data is mostly noise. Defense context still appears in each pick's reason.
• Regular season only.
• GitHub Pages lets browsers cache pages for 10 minutes, so hard refresh (⌘⇧R) right after a deploy. Scripts and styles are cache-busted (see above). The HTML itself can still be up to 10 minutes stale.

Ideas and next steps are in ENHANCEMENTS.md.