DGPT Standings Forecast

Live: dgoodenough.github.io/discgolf

I built this to answer one question all season: who makes the Powerball Cup? It simulates the rest of the 2026 Disc Golf Pro Tour 100,000 times and publishes qualification odds for every player, updated within 15 minutes during events. Points are explained for newcomers at how it works.

I first built a version of this in 2021 as a Jupyter notebook fed by hand-pulled CSVs (preserved in archive/2021). This is the rebuild: PDGA API data, exact points rules, and a site that maintains itself.

What it does

• Standings, computed independently. Results come from PDGA live scoring, points from my implementation of the 2026 rules: per-place curves, class multipliers, tie splitting, per-class counting caps, winner invites, the doubles team curve, Jomez bonuses. Totals match StatMando, the official administrator, exactly. A CI check re-verifies that after every weekly refresh and fails loudly if we ever drift.
• A calibrated simulation. Scores are drawn from PDGA ratings. I refit the model against all 6,667 completed 2026 player-rounds: about 6 rating points equal one stroke per round in MPO (7.3 in FPO), with a 4.2-stroke event-level round spread. Fields use real registration lists where they exist and cohort-based participation rates (tour card, European) where they don't. The playoffs read their published signup lists and take them literally — an entered player is in that field however the standings finish — and fall back to the qualification ladder only for players whose registration wave hasn't opened. Worlds includes its one-round play-in: 6 MPO / 2 FPO spots, raced against the real entry list.
• Live events, modeled mid-round. During play, each player's current score is locked in and only their remaining holes are simulated, so odds track the actual tournament instead of resetting to priors.
• The live race, kept as a series. The simulation answers "who wins this tournament" every time it runs; those answers are now recorded, so the Event odds tab draws every contender's win probability through the event — including the players who led it and lost it — over a table of everyone still above 0.1%. Under it, the fork line: how far back the tournament is still live, drawn at three tolerances. Each line is the score the eventual winner beats 4 times in 5, 19 in 20, or 199 in 200, so the label on a line is how often calling everyone below it done is wrong. Both charts run on a clock with the hours nobody was playing cut out of it, or on the field's own progress through the holes, and either crops to one day's play. Under the axis, a tally of skulls — one per player who, at that moment, took none of the ten thousand simulated tournaments — and a phoenix for anyone who climbs back out. The table lists the whole field with the moment each one went.
• The seed, not just the spot. The Cup starts every qualifier on a score set by their World Standings position — -7 for the MPO No. 1, -6 for FPO, down to even for the bottom seeds. The Starting strokes column runs every simulated season through that ladder, so one row carries both the score a player would tee off on and how often they get no tee time at all.
• The championship, played out. Once the seed is known the Cup is a four-round tournament like any other, so the model runs it: scores from ratings, the seed's head start added to the total, low score wins. The Win Cup column is the result, and unlike qualification odds it is a race — the field's numbers add to 100%.
• A what-if for every player. Expand any row, toggle which events they will attend, and their Cup odds recompute in under 100ms against 25,000 frozen simulation cutlines.
• Its own scorecard. Every meaningful forecast is snapshotted to predictions/ so the model can be graded (Brier, calibration) at season's end with python -m dgpt.evaluate.
• Next season, clearly labelled as a guess. The DGPT has published its 2027 schedule, so two experimental tabs project seasons nobody has played: the whole 2027 DGPT season, and the new EuroTour. Both are built from the announcement plus this year's ratings and attendance rates; see what they assume below.
• And one for a player who doesn't exist. The What if tab drops an invented player into the projected 2027 field: set a rating and how many DGPT, DGPT+, major and JomezPro events they enter, and every column the 2027 table carries is recomputed in the browser, 8,000 simulated seasons per slider move. The playoffs are not sliders — those are earned on points, which is half of what is being computed.

The experimental tabs

dgpt/project.py forecasts a season that has not started. It is deliberately separate from simulate.py — the live 2026 forecast is the product and updates every fifteen minutes during play, so it should not grow a "no results yet" mode — but the two share what must not fork: the score model, the points curves, the participation model and the ranking helpers.

Every assumption lives in dgpt/season2027.py, is emitted with the data, and is shown on the page above the numbers. In short:

• The calendar is real. data/schedule_2027.csv is the announcement, transcribed. Round counts come from the dates (three-day stop, three rounds), which is the rule every 2026 event obeys. Event ids are synthetic — the PDGA has not assigned 2027 ones — so nothing links out to a PDGA event page yet.
• 2027 DGPT assumes 2026's rules. Same curves, class multipliers, per-class counting caps, playoff windows and Cup seed ladder, with Ivy Hill inheriting the Green Mountain Championship's qualification and the Kansas City Wide Open the MVP Open's. The announcement changes the calendar and the postseason cadence, not the points structure. What is genuinely new is data: the USDGC joins the MPO race as a fourth major, and the JomezPro Series triples to seven all-counting stops.
• Players carry today's ratings and this year's attendance rate. Each is projected to play the same share of US stops, European stops and JomezPro stops as in 2026 — on a longer calendar, so more starts. Nobody improves, retires or turns pro, and a 2027 rookie does not exist.
• The EuroTour's points structure is the published one. Three categories paying 250 / 200 / 150 for a win, counted best 1-of-2, 2-of-3 and 3-of-3 — six results, and 1,100 for a perfect season, which season2027.max_points() asserts. The 2028 card bands (MPO top 6 Full / top 24 EuroTour, FPO top 3 / top 12) are theirs too. Ours is only the shape of each curve below first place: the DGPT's own per-place curve scaled to the published win value, which lands category 3 on exactly the Elite Series curve and 2 and 1 on the DGPT+ and Playoff multipliers.
• What the EuroTour tab is still missing is the field, not the rules. Its roster is European players with a 2026 DGPT standings row, so domestic-only Europeans are absent from a race they would really be in. At the three stops that are open to the whole tour — the European Open and both European Elite Series events — the table is at least raced against an outside field drawn from the rest of the standings, so a European's finish there is not measured against Europeans alone. Displacement (a Full Tour Card holder passing their EuroTour Card down) is not modelled and only ever pushes cards deeper, so the published odds are a floor for anyone just outside a band.

Both run at 20,000 simulated seasons in the daily refresh (--project-sims, or --skip-projection to leave them alone, which is what the live loop does).

The What-if tab

dgpt/whatif.py rides along with the 2027 DGPT projection and reduces it to a ~25 KB summary — the season total sitting at each standings position as a quantile ladder, both playoff qualification lines, the odds a given finish at a playoff event wins a Cup place, the ratings behind each Cup seed, and each event's field as twenty equal-weight rating bands. dropIn() in docs/js/sim.js draws seasons against it in the browser.

Four approximations are the tab's own, on top of everything the 2027 projection already assumes, and the page lists all four:

• The field is frozen. A dropped-in player is added to the standings, not swapped into them, so the points they win are not taken off anybody. That flatters them most where the race is tightest — and at the very top it means racing a player the model still has winning the season. Correcting it needs a re-simulation of the field, which a slider cannot do between frames; it is the same trade the 2026 row expander's what-if makes.
• Which events is a draw, not a plan. Each simulated season picks a fresh random subset of the size asked for, because nothing here knows which stops a given player would choose.
• The doubles championship is left out. It needs a partner, and 2027 pairings cannot be known.
• The Cup's opponents are averages — the mean rating at each seed rather than a distribution over who holds it.

The ladder is shipped as marginal quantiles per rank and read at one uniform, which is a comonotone approximation of the joint distribution. It is exact for every number the tab prints, because each is a statement about a single rank: P(rank <= r) is P(total > the r-th best total), and that rank keeps its true marginal. dgpt/whatif.py's docstring is the full argument.

Running it

pip install -r requirements.txt
cp .env.example .env        # PDGA API credentials (developer program)
python -m dgpt.refresh --sims 100000

That rebuilds the schedule, banks any newly finished events, re-runs the simulation, and regenerates the site data in docs/. GitHub Actions runs it twice weekly, plus every 15 minutes during live play (a cheap change-check skips the heavy work between rounds). python -m dgpt.validate diffs the standings against StatMando; python -m dgpt.calibrate refits the score model from cached rounds.

Data sources

• PDGA REST API (schedule, ratings; requires developer credentials), PDGA live scoring (results, registrations), and PDGA event pages (signup lists for the playoffs and the Worlds play-in, which live scoring doesn't carry until event week). Event and player data © 2026 PDGA. PDGA Authorized Developer.
• DGPT points structure and playoff qualification
• The DGPT's 2027 schedule announcement (Elite Series, JomezPro Series and EuroTour), transcribed by hand into data/schedule_2027.csv — it is the only source the experimental tabs have
• StatMando for validation only

Styled with Ledger. Inspired by FiveThirtyEight's sports forecasts.

License

MIT.