How It Works

USR (Universal SciOly Rating) is an experimental rating built only from official event placements in public tournament results. It rates each event separately, adjusts for how strong each event's field was, and then combines events into an overall number. Model version v2-exp.2.

In plain language

Every event is its own contest. For each tournament and event, we take the official place among the teams that actually received a comparable placement. No-shows, participation-only results, disqualifications, and exhibition entries are shown in results but are not counted in the rating. Missing data is treated as unknown, never as zero.
Places become scores. Finishing r-th of n becomes x = ln((n + 1 − r) / r). The middle finisher scores 0, winning a bigger field scores higher, and ties share the average place.
Field strength is estimated, not assumed. Beating strong teams is worth more. Every tournament event gets an estimated field strength k, fitted jointly with every team's event skill, so a team's rating depends on who it actually beat.
Recent and larger fields count more. Results fade with a 200-day decay, only the last 400 days count, and bigger fields carry a little more weight.
Thin evidence is pulled toward average. A team with one result is shrunk toward the middle (USR 10) much more than a team with ten results.
Overall combines all official events. An official event with no comparable result counts as average (the prior), so coverage is always shown next to the number. Official total points and placements are kept separately; overall placement is not counted twice.
Two views that never mix. Team Performance rates actual team entries (e.g. “Gold”) within one season. School Potential superscores a school's entries event by event within each tournament and re-ranks unique schools — useful for understanding depth, but it may not match any roster a school could field. Division B and C are separate pools.

Eligibility, status, and team identity

Established (numbered national rank): nationally comparable results in all M official events of the season, from at least 3 distinct tournaments, with a result in the last 180 days. M comes from the season's official event list (23 in both divisions for 2025–26), not a hard-coded number.
Provisional: rated, but with missing or local-only events or too few tournaments. Shown with exact coverage and no national rank.
Inactive: no eligible result in 180 days. A profile with no current-season result is not in current-season rankings.
Ranks and filters. National and state ranks are computed on the full established pool before any filter or page is applied; filtering never renumbers them.
Team identity. A school is matched on name, city, and state; same-named schools in different places are never merged automatically. A team is school + division + season + label. Team numbers are local to a tournament and never used as identity. An unlabeled entry is its own “unlabeled team” only when it was the school's only entry at that tournament; otherwise it is unresolved and excluded from Team Performance until a reviewed mapping resolves it. Unlabeled teams are never assumed to be “A”.
Season transitions. Team-season identities restart each season (rosters change). School Potential may use the previous season within the 400-day window, but only for official events explicitly listed as equivalent; a rotated-out event never carries into its replacement.
Trial events. A trial or trialed event counts only if it is one of the season's official national events with valid placements. Other trial events are displayed but not rated.
Sparse and disconnected data. For every event we build a graph of teams and tournament fields. Only the largest connected component is nationally comparable; teams in smaller components get local-only event ratings that do not enter national overall ranks.

Worked example (fictional teams)

Computed live by the production rating code. The teams and tournaments below do not exist.

One event, School Potential view, as of 2026-02-15. Two fictional tournaments:

Tournament	Team	Place / n	x = ln((n+1−r)/r)	Weight (normalized)
Harbor Invitational (fictional)	Tern Lake A	1 / 4	1.386	0.971
Harbor Invitational (fictional)	Quill Ridge	2 / 4	0.405	0.971
Harbor Invitational (fictional)	Basalt Park	3 / 4	-0.405	0.971
Harbor Invitational (fictional)	Juniper Flats	4 / 4	-1.386	0.971
Summit Regional (fictional)	Quill Ridge	1 / 3	1.099	1.039
Summit Regional (fictional)	Juniper Flats	2 / 3	0.000	1.039
Summit Regional (fictional)	Oxbow Prep	3 / 3	-1.099	1.039

The fit estimates field strengths: Harbor k = 0.130, Summit k = -0.149 (positive = stronger field). Adjusted performance is x + k.

Team	Skill s	Potential q	Event USR (from q)	Shrinkage
Tern Lake A	0.747	0.837	12.06	51%
Quill Ridge	0.500	0.529	11.32	33%
Basalt Park	-0.136	-0.131	9.67	51%
Juniper Flats	-0.457	-0.369	9.08	33%
Oxbow Prep	-0.636	-0.515	8.72	49%

Quill Ridge finished 2nd at Harbor but won Summit and beat Juniper Flats twice; Tern Lake A's single win is shrunk more heavily because it is one result. Oxbow Prep's only result is last place in the weaker Summit field.

Does it work? Backtest results

Chronological by tournament: each tournament is predicted only from refits dated before it began. Pairwise accuracy = share of correctly ordered pairs among entrants every model could rate.

Backtest results have not been generated for this database (run npm run backtest).

Credit and what is adapted

The event-wise approach is informed by SentientTree's 2026 SO Rankings (FAQ tab and the Division B overall formula, inspected 2026-09-27): placement logits, superscoring, field (competitiveness) adjustment, recency and field-size weights, and a softplus generalized-mean aggregate. scly.io does not claim to reproduce those rankings. The following are scly.io adaptations:

A specified regularized least-squares fit for skills and per-event field offsets (the reference's iteration details are unpublished).
Recency uses age before the as-of date (the reference's sign convention is unclear); no competitiveness multiplier in weights.
A prior term (λs·softplus(0)) in School Potential summaries, and missing events count as the prior rather than contributing nothing.
No discretionary exclusions; no overall/event blending for sparse fields (explicit components and evidence labels instead).
A separate Team Performance view; the reference superscores all teams.
No predicted placements or win probabilities until their calibration has been tested.

UTR Sports inspired the product idea of a searchable, explainable rating. scly.io is not affiliated with UTR and does not use its algorithm.

Technical details

Show formulas and parameters

Observation. For event e at tournament t with n eligible participants (entries in Team Performance, unique schools in School Potential) and model rank r (midranks for ties): x = ln((n+1−r)/r). Events with n < 2 produce no observation.

Weight. w = exp(−age/200) × N^0.25 × format, where age = as-of − end date (days, within a 400-day window, never after the as-of date), N = unique eligible schools, format = 1 in person, 0.5 explicitly online, 1 unknown (flagged). Weights are normalized to mean 1 in each event pool.

Fit. Minimize Σ w (s_i − k_t − x_it)² + λs Σ s_i² + λk Σ k_t² with λs = 1, λk = 1. This is strictly convex with a unique solution, solved with preconditioned conjugate gradient and accepted only when one alternating update moves no parameter by more than 1e-7 (cap 10,000 iterations). Non-converged fits block publication.

Event values. Team Performance uses s. School Potential uses q = softplus⁻¹((Σ w·softplus(x + k) + λs·softplus(0)) / (Σ w + λs)).

Overall. Team Performance: mean of s over all M official events (0 for missing or local-only). School Potential: softplus⁻¹(mean of softplus(q)) over M events (q = 0 for missing).

Display. USR = 20 / (1 + exp(−z / 2)): z = 0 → 10. Two decimals shown; sorting uses full precision. The scale is a presentation convention; values are not calibrated across divisions, views, seasons, events, or disconnected components.

Refits and history. Ratings are refit every Sunday from results completed by that date. Change explanations decompose each move exactly into new results, recency, field recalibration, and window/coverage. Pre-tournament field strength uses the last refit strictly before the start date; per-event k values shown on tournament pages are retrospective.

Evidence labels (heuristics, not confidence intervals): strong = 3+ appearances and n_eff ≥ 2.5; moderate = 2+ appearances; limited = 1; local only = outside the reference component. n_eff = (Σw)²/Σw².

Data sources and coverage limits: Data coverage.