How the forecasts work, and how accurate they are
Threshold Ladder forecasts how many points a given rank will take in a Fortnite tournament, before it starts and while it runs. This page explains where the numbers come from, how they are made, how they are tested, and where they should not be trusted.
The data
Everything starts from Osirion's public Fortnite API: Epic's tournament calendar, each event's scoring rules, and the final leaderboards. A harvester reads them every day; the model is trained on 99,497 thresholds from 8,116 tournaments (model of 23 October 2026). A threshold is the points held by the team at a given rank when the leaderboard is final.
The leaderboards also record, for each team, its placement and eliminations in every game and when each game ended. That makes it possible to rebuild the board at any moment of a past session, which is how the live part of the forecast is measured, and to score past games again under another table, which is how new cups are priced.
What the site publishes is derived: forecasts, and statistics such as medians and shares. It does not republish leaderboards; the full standings of a cup are one click away on osirion.gg, which the forecast links to.
Before the cup: a cascade of readings
The forecast tries the most direct reading first and moves down only when it has to:
- The cup's previous editions. The same cup, region, format and rank, the last times it was played the same way, the latest counting most. Corrected for a field that differs from last time, for a change in who may enter, and for a change of season.
- The cup's level times its measured shape, for ranks its editions did not reach: the threshold at rank 20, times what each rank was worth against rank 20 in its editions.
- The ladder, for ranks no edition measured: what each rank is worth against rank 20 across every open queue of the same size and mode (see deep ranks).
- Recent leaderboards replayed under the cup's table, for a cup with no edition in its region (see brand-new cups), combined with its editions in other regions where it has some.
- The finals of the same format, for a final played in a single lobby.
- The scoring table alone, as the last resort: the share of the maximum possible score that rank 20 reaches in cups of the same kind.
Every forecast carries a range. The half-width comes from how much the reading used has moved in the past – between two editions of the cup, for the first rung – and how wide the range must be for half, or nine in ten, of the final thresholds to fall inside it is measured on tournaments the model had not seen.
During the cup: the live board
While a cup runs, a small service reads its public standings every five minutes. Each reading is compared with the share of its final value a threshold has typically reached at that point of the session – measured on 3,690 replayed open queues and 655 single lobbies, by kind of cup – and weighed against the forecast made before the cup, each according to how precise it has proven on past evenings. Towards the end of the session the history's weight shrinks further as the cup advances, by how much is learned on past cups, and a rank the board does not show follows the nearest rank it does show. The points already on the board are a floor: a threshold never falls. How the cutoff moves during a cup goes into detail.
How accurate it is
The test that matters is the one a player lives: forecasting a cup from what had finished before it. The newest 600 tournaments – everything since 15 September 2026 – were each forecast only from the tournaments that had finished before their day, and compared with their final leaderboards (6,409 thresholds):
| What is forecast | Median error | Inside the quoted range |
|---|---|---|
| Every threshold of the newest 600 tournaments | 6.4% | 93% |
| Cups with a previous edition (where the model and both baselines answer) | 3.9% | — |
| … the same rows, last edition carried forward | 4.2% | — |
| … the same rows, median of the cup's category | 5.3% | — |
| Cups never seen in their region, from the scoring table alone | 9.5% | — |
| Finals played in a single lobby | 36.3% | — |
Median absolute error, in per cent of the final threshold. "Inside the quoted range": the share of final thresholds inside the forecast's main range, which claims 80%; the site's own 50% and 90% ranges are measured from the same errors. Measured 22 October 2026, and measured again every three days as tournaments are added.
Three honest caveats. The cups the model has seen before are the easy ones, and most of the error is concentrated in the others. The top five ranks are the noisiest part of any leaderboard: there, carrying last week forward is about as good as the model. And a single evening can always land outside the range – one in ten, by construction, lands outside the 90% range.
Where it stops being reliable
- The first cups of a new season, before anything has shown which way the season moves thresholds (see new seasons).
- Formats with no history and nothing to replay: the very first cups of a new mode, Arena-style events with no fixed number of games.
- Very deep ranks in the biggest fields: the public leaderboards stop at about 9,950 teams, so ranks near that ceiling rest on fewer leaderboards.
- Anything the rules change on the day – a session shortened, a scoring table corrected after publication. The forecast reads the rules as published.
The code and the full notes
The whole project is open source. The page and its forecasting code are in the threshold-ladder repository; the harvester, the model's training and every measurement quoted on this site are in fortnite-tracker, whose methodology note records each change to the model and what it was measured to do.
Updated 23 October 2026