Methodologyv2.0

Every probability is produced by the Spolytics Engine. This page states what data the engine uses, what it adjusts for, and the tests a number must pass before it appears. The internal formulation is not published; the calibration coefficients and scores running on the server right now are.

Probability

Per-league strength models compute a raw probability via scoreline grids (soccer), per-PA simulation (baseball) and strength index (esports).

Interval

The raw number passes a calibration function fitted on finished games, and ships with a 95% interval. Thin samples are labelled provisional.

Per-sport record

Graded daily on Brier, ECE and RPS; only coefficients that win are kept. Live coefficients and pre/post-fit scores are in the table below and on /report.

See the accuracy dashboard →

What changed in v2.0

Soccer became the core. Spolytics Engine parameters are fitted per league and refit weekly in-app, and totals, BTTS and margin lines all come off the same scoreline grid.
A shot-level simulation simulator was attached in shadow mode. It runs and is graded daily, but its blend weight is 0, so displayed probabilities are unchanged for now.
Player Probabilities (anytime scorer, shots and shots on target) were added on top of the shadow simulator. They are also graded daily but stay at weight 0 and off the display.
All pitcher strikeout markets were removed. Baseball keeps result, totals and batter markets only.
NFL added a starting-QB EPA adjustment (±120 rating points); esports added map-level strength index turned into series probabilities.
A guard was added to live probabilities: an implausible value is withheld and labelled unavailable rather than shown.
Leagues that are not fully modelled were removed from the board.

Spolytics Engine

v1.2

Every probability here is computed by our own engine, from event data rather than repackaged odds or someone else's projections. v2.0 makes soccer the core with per-league models refit weekly, and new simulators are validated in shadow before they count.

It does not start from odds
Most published probabilities are reverse-engineered from market prices. The engine never takes odds as an input; every number is computed from event data alone.
Markets cannot contradict each other
From v2.0, soccer result, totals, BTTS and margin all come off one scoreline distribution. Nothing is computed market by market, so two numbers cannot disagree.
New models run in shadow first
A new engine such as the shot-level simulation runs and is graded daily, but stays at weight 0 with no effect on screen until it beats the existing model.
Probabilities are locked at publication
The first value shown becomes the graded record. A later model change never rewrites a past prediction.
It re-grades itself every morning
Every finished game is scored and the calibration coefficients are refit. Losing predictions stay in the sample.
Shadow Player Probabilities are validated too
After v2.0, soccer anytime scorer, shot and shots-on-target props are also computed and graded daily on the shot-level simulator. They also run in shadow at weight 0 and do not appear until they beat the incumbent.
No single signal can flip a call
Injuries, form, xG, park and umpire adjustments are each capped at ±4-6% and thin samples are pulled toward the mean. An implausible live value is withheld rather than shown.

Read the full methodology →Model updates →Last change 2026-09-12

The computation pipeline

The seven stages a single probability passes through before it appears. Yesterday's grading feeds back into the calibration stage each morning.

01Collect
Schedules, lineups and event logs are pulled from public APIs, refreshed per game because each league lands at a different time.
02Context
Park, weather, umpire, travel and rest — plus soccer injuries, lineups and recent xG — enter as separate coefficients, each capped at ±4-6%.
03Compute
Soccer builds a scoreline distribution from a per-league fitted model and reads result, totals, BTTS and margin off that same distribution. Other sports combine team and player strengths on a head-to-head scale. Odds are not an input.
04Guard
Thin samples are pulled toward the mean and bounds tighten with competition reliability, so no single signal can flip a call.
05Calibrate
Coefficients fitted on past predictions versus actual results are applied, so games called at 80% land near 80% in reality.
06Shadow test
A new engine such as the shot-level simulation computes the same games in parallel and is graded daily. Its weight stays at 0 until it beats the incumbent.
07Lock & publish
The first published value becomes the graded record; a later model change never rewrites it.

Feedback loop — every finished game is graded each morning and the error refits the stage 05 calibration coefficients. Losing predictions stay in the sample.

Computation rules by version

What changed, rule by rule, between engine versions. A rule change never recomputes probabilities already published under an earlier version.

Rule	v1.0 · Previous · 2026-08	v2.0 · Current · 2026-08
Soccer core model	Team strengths to a scoreline grid with a low-score correction	Spolytics Engine parameters fitted per league and refit weekly inside the app
Soccer shot-level simulation	None	A 20,000-run shot-level simulation runs daily in shadow; with the blend weight at 0 it does not move a displayed probability
Soccer shadow Player Probabilities	None	Anytime scorer, shots 1+/2+ and SOT 1+ are computed and graded daily on the shot-level simulation; with soccer.simProps.enabled at 0 they do not affect the display
Soccer market coverage	Mostly win / draw / loss	Totals, BTTS and margin lines (-1.5 etc.) derived from the same scoreline grid, each with a written rationale
EPL data	Public schedule and result feeds	Shots, form, injuries and lineups ingested from a paid feed into our own store and fed into xG context (capped at ±4%, shrinkage k=6)
Pitcher strikeout markets	MLB pitcher K probabilities published	Removed. Baseball keeps result, totals and batter markets only
NFL team rating	Team-level metrics only	Starter's EPA per dropback compared with the team baseline, capped at ±120 rating points
Esports	None	Map-level strength index turned into LoL and CS2 series probabilities (Bo3/Bo5 binomial), major leagues only
Live probability guard	Computed value displayed as-is	If a team trailing badly near the end reads above 25%, the number is withheld and labelled unavailable
League exposure rule	Fixtures shown even without a probability	Only fully modelled leagues appear on the board; the rest are listed with status on /coverage
Calibration refit	Automatic each morning after grading every finished game	Morning grading plus a 21-day rolling refit, with the constant registry evolving nightly within a per-run step cap
Probability locking(unchanged)	First published value fixed as the graded record	First published value fixed as the graded record
Grading sample(unchanged)	Every finished game, nothing excluded	Every finished game, nothing excluded

Soccer core model
v1.0
Team strengths to a scoreline grid with a low-score correction
v2.0
Spolytics Engine parameters fitted per league and refit weekly inside the app
Soccer shot-level simulation
v1.0
None
v2.0
A 20,000-run shot-level simulation runs daily in shadow; with the blend weight at 0 it does not move a displayed probability
Soccer shadow Player Probabilities
v1.0
None
v2.0
Anytime scorer, shots 1+/2+ and SOT 1+ are computed and graded daily on the shot-level simulation; with soccer.simProps.enabled at 0 they do not affect the display
Soccer market coverage
v1.0
Mostly win / draw / loss
v2.0
Totals, BTTS and margin lines (-1.5 etc.) derived from the same scoreline grid, each with a written rationale
EPL data
v1.0
Public schedule and result feeds
v2.0
Shots, form, injuries and lineups ingested from a paid feed into our own store and fed into xG context (capped at ±4%, shrinkage k=6)
Pitcher strikeout markets
v1.0
MLB pitcher K probabilities published
v2.0
Removed. Baseball keeps result, totals and batter markets only
NFL team rating
v1.0
Team-level metrics only
v2.0
Starter's EPA per dropback compared with the team baseline, capped at ±120 rating points
Esports
v1.0
None
v2.0
Map-level strength index turned into LoL and CS2 series probabilities (Bo3/Bo5 binomial), major leagues only
Live probability guard
v1.0
Computed value displayed as-is
v2.0
If a team trailing badly near the end reads above 25%, the number is withheld and labelled unavailable
League exposure rule
v1.0
Fixtures shown even without a probability
v2.0
Only fully modelled leagues appear on the board; the rest are listed with status on /coverage
Calibration refit
v1.0
Automatic each morning after grading every finished game
v2.0
Morning grading plus a 21-day rolling refit, with the constant registry evolving nightly within a per-run step cap
Probability locking
v1.0
First published value fixed as the graded record
v2.0
First published value fixed as the graded record
Grading sample
v1.0
Every finished game, nothing excluded
v2.0
Every finished game, nothing excluded

See the full update log →

Anti-tout declaration

We do not handle odds, lines or sportsbooks. No odds exist anywhere on this site, and we never compute an edge against them.
We never produce a number like "90% hit rate". The only figures we publish are Brier score, log loss, calibration error (ECE), and RPS for soccer.
Published probabilities are locked. The first published value becomes the graded record, and a later model change never rewrites it.
Losses are published with the wins. Grading covers every finished game and we never drop a failed prediction from the sample.
Thin samples are labelled provisional. Brier is unstable at low match counts, so we do not dress it up as a settled track record.

1. Calculation methodology

Every probability is computed by our own engine — a simulation and statistical framework built specifically for this purpose, not a repackaging of odds or someone else's projections. We don't publish the specific mix of data providers behind it, but every number is produced by our own calculation method, described in the sections below (simulation, calibration, walk-forward validation).

2. How the engine computes

The engine's internal formulation is proprietary and not published. What moves a probability in each market is stated in full below.

Soccer

Attack and defence strengths produce home and away scoring rates, probability is spread across the full scoreline grid, and the low-scoring cells (0-0, 1-0, 0-1, 1-1) are lifted to match observed frequency. xG is our own shot-location and shot-type proxy, not a licensed data product. Home advantage and the size of that lift are fitted and stored per league; a league without a fit runs on defaults and is labelled as such.

Win probability

Both teams' strengths are combined on a head-to-head scale with home advantage added. The strength index keeps updating from recently completed games.

Soccer markets

Totals (over/under), both-teams-to-score and margin lines (-1.5 and similar) are not computed separately from the match result. Each is the sum of the relevant cells in the same scoreline grid, so no two markets can contradict each other. Every market carries a written statement of what drove it.

Soccer shadow simulator

v2.0 adds a shot-level simulation (20,000 replays per match). It generates shots and per-shot conversion to build a scoreline distribution. It currently runs in shadow: results are stored daily and graded against real outcomes, but the blend weight is 0, so displayed probabilities are unchanged. The weight rises only once grading shows it beats the existing model.

Soccer shadow Player Probabilities

Anytime scorer, shots 1+/2+ and SOT 1+ are broken down player-by-player in the same shot-level simulator. Each player's last 15-match shot share, SOT rate and finishing are shrunk k=6 toward league/team averages, and playing time is scaled by start probability. While soccer.simProps.enabled is 0, they are only stored and graded and do not affect the display.

EPL ingestion

For the EPL we ingest matches, shots, team form, injuries and lineups from a paid feed into our own store. The resulting xG context enters behind a ±4% cap and shrinkage (k=6), so a single match cannot swing a probability. Absences are weighted by the player's share of team goals rather than headcount (50% bench compensation, capped at 18%).

NFL

Team ratings carry a starting-quarterback adjustment: the starter's EPA per dropback against the team baseline, moving the rating by at most ±120 points. It captures the gap that team-level metrics miss when a backup starts or the job changes mid-season.

Esports

LoL and CS2 run on map-level strength index. Series probabilities (Bo3/Bo5) are the binomial sum over map win rates. Minor circuits with thin samples are not shown at all.

Live probabilities

In-play probabilities are computed and labelled separately from pre-game ones. If less than 15% of game time remains and a team trailing significantly reads above 25%, the value is discarded and shown as unavailable. Withholding a number beats displaying a plausible-looking wrong one.

Batter markets

Computed per plate appearance. PAs in one game share a pitcher, a park and an umpire, so they are treated as correlated rather than independent. Opposing starter and bullpen, platoon splits, park factors and a 15-game form window feed the per-PA rate.

Simulation

Combinations with no closed form (usage that depends on game flow, correlated props) are simulated. Simulations always retain the distribution, and the percentage shown is a probability read off that distribution.

3. Calibration layer

A raw model probability is never displayed directly. It passes through a calibration function fitted on (raw probability, actual outcome) samples from finished games. Calibration preserves ranking and only corrects the probability, so adopting it strictly on hold-out improvement carries no downside.

Fitting unit

Market family (Win Probability, draw, pitcher Ks, batter power, ...) x scope (league -> sport -> global). Lookup tries the league-specific coefficient first, then sport, then global. Global coefficients mix base rates across sports, so they require a far larger sample (n >= 1000).

Method selection

A high-degree-of-freedom stepwise correction is only a candidate at n >= 1000, because it overfits on small samples. Below that, only gentle curve and slope corrections are used — forms that cannot damage an already well-calibrated model.

Adoption test

A candidate is adopted only if it lowers hold-out Brier by at least 0.0005 across a 4-fold walk-forward validation, refitting inside each training window so no future data leaks in. Families that fail the test ship uncalibrated.

Guard rails

Minimum sample n >= 150 (1000 for global), a calibration slope floor of 0.35 (below that probabilities collapse to a point and information is lost), and a hard shift cap of +/-12pp. Calibration is a nudge; anything larger is clipped.

Refitting

Refit on a 21-day rolling window. Rosters, pitcher usage and umpire zones drift through a season and the bias drifts with them, so no coefficient is hard-coded permanently. Coefficients that lose their evidence in the current window are deleted.

Stacking

Groups with n >= 400 also test a blended correction that re-weights several signals. The baseline is the already-calibrated probability, so the blend ships only if it beats that; otherwise calibration alone stands.

4. Calibrators running right now

This table is not documentation filler: it is the set of coefficients the server used for this request. It changes every time the refit job runs. Last fit: 2026-10-03 13:11 UTC

Scope	Family	Method	n	Brier before	After
LoL	Win Probability	Slope	182	0.2177	0.1535
ESPORTS	Team totals	Slope	435	0.2784	0.2422
CS2	Team totals	Curve	344	0.2734	0.2431
MLB	Batter contact	Stepwise	2,946	0.2198	0.2027
BASEBALL	Batter contact	Stepwise	2,946	0.2198	0.2027
Global	Batter contact	Stepwise	2,946	0.2198	0.2027
Global	RBI / runs	Slope	2,946	0.2159	0.2035
MLB	RBI / runs	Slope	2,946	0.2159	0.2035
BASEBALL	RBI / runs	Slope	2,946	0.2159	0.2035
ESPORTS	Totals	Slope	435	0.2452	0.2351
J League	Soccer goals	Curve	180	0.1319	0.1229
CS2	Totals	Slope	344	0.2480	0.2409
Club Friendly	Soccer goals	Curve	920	0.2429	0.2362
BASEBALL	Walks	Curve	1,473	0.1917	0.1860
Global	Walks	Curve	1,473	0.1917	0.1860
MLB	Walks	Curve	1,473	0.1917	0.1860
BASEBALL	Pitcher Ks	Slope	440	0.2045	0.1999
MLB	Pitcher Ks	Slope	440	0.2045	0.1999
MLB	Batter power	Curve	2,946	0.1568	0.1527
BASEBALL	Batter power	Curve	2,946	0.1568	0.1527
Global	Batter power	Curve	2,946	0.1568	0.1527
LaLiga	Soccer goals	Curve	210	0.2411	0.2406

5. How we grade ourselves

Brier

Mean of (probability - outcome)^2. Lower is better; 0.25 is the always-50% baseline. A model that cannot beat that baseline carries no information.

Calibration curve / ECE

Bucketed check that predictions stated at 70% actually land 70% of the time. ECE is the weighted average of that gap.

RPS

Ranked probability score, used where the outcome is ordered (win/draw/loss). It penalises a model that systematically misses draws.

What we compare against

We benchmark our model against public crowd forecasts only. Closing lines and odds-implied probabilities are not used even as a comparison, because we keep odds off this site entirely. That is also why no betting metric such as CLV appears anywhere here.

See the accuracy dashboard · Validation report · Daily learning log

6. Known limits

Match-result accuracy in soccer sits structurally around 49-58%. That is the ceiling for any statistical model of this kind; anyone promising more is selecting their sample.
A league that is not fully modelled does not appear on the board at all. Mixing fixture-only leagues in makes it impossible to tell where the computed numbers stop. Per-league status is listed in the coverage document.
Calibration does not always win. Some families show no hold-out improvement and ship raw: if a family is missing from the table above, it is uncalibrated.
Injury and lineup news changes right up to first pitch. A probability reflects the information available when it was locked, and nothing after that.

Spolytics is a statistical research tool. Nothing here is a betting recommendation, and we provide no odds, lines or sportsbook information.