The Model
FPL Predictor never forecasts points from past points. It models the football events that create FPL points and converts those probabilities into expected points.
2026/27 FPL scoring rules used by every projection
| Action | GK | DEF | MID | FWD |
|---|---|---|---|---|
| Plays 1–59 mins | 1 | 1 | 1 | 1 |
| Plays 60+ mins | 2 | 2 | 2 | 2 |
| Goal | 10 | 6 | 5 | 4 |
| Assist | 3 | 3 | 3 | 3 |
| Clean sheet | 4 | 4 | 1 | 0 |
| Every 3 saves | +1 | — | — | — |
| Penalty save | +5 | — | — | — |
| Defensive contributions | — | +2 at 10 | +2 at 12 | +2 at 12 |
| Bonus points | 1–3 | 1–3 | 1–3 | 1–3 |
| Penalty miss | -2 | -2 | -2 | -2 |
| Yellow card | -1 | -1 | -1 | -1 |
| Red card | -3 | -3 | -3 | -3 |
| Own goal | -2 | -2 | -2 | -2 |
| Every 2 goals conceded | -1 | -1 | — | — |
Defensive contributions are capped at 2 points per match: defenders need 10+ clearances, blocks, interceptions and tackles; midfielders and forwards need 12+ of the same plus ball recoveries. Bonus points come from the BPS ranking (1st +3, 2nd +2, 3rd +1).
1 · Market prior (season odds)
The latest season outright prices — title, top four, top half, bottom half, relegation and finish bottom — are converted to implied probabilities, de-vigged so each market sums to the number of places it pays, then blended into a single team quality index. Markets price transfers, managers and squad depth that FPL's own strength numbers miss, so they carry 65% of the pre-season prior.
2 · Team strength (Elo)
Each club carries separate attack and defence ratings on a log scale, seeded from the market index plus FPL strength priors and updated after every completed fixture. Updates are weighted by opponent quality and recency, so beating a strong side moves the rating more.
3 · Fixture model (Poisson)
Every upcoming match is turned into expected home and away goals, then a Poisson score distribution gives win/draw/loss and clean-sheet probabilities. Our own 1–5 difficulty comes from this, not from FPL's FDR.
4 · Injury news & predicted XIs (scraped)
We scrape live injury tables and predicted line-up articles every 45 minutes, use an LLM to normalise them into structured entries, and resolve the names onto FPL players. Ruled-out players are capped near zero availability, doubts are halved, players named in predicted XIs get a start-probability floor, and anyone predicted to be benched gets capped. Every flag is visible on the Team news page.
5 · Expected minutes
Start rate is shrunk toward a blended prior — last completed season's start share (weighted by how many minutes the player actually banked) mixed with a price-implied baseline — then adjusted for FPL injury/suspension status, chance-of-playing flags and the scraped team-news overrides above. Outputs: P(start), P(60+), P(any minutes) and expected minutes — these scale everything else.
6 · Attacking events
xG90 and xA90 are Bayesian-shrunk toward a personalised prior (500 pseudo-minutes, tightened to ~420 in the opening weeks): last completed season's own per-90 record blended with a price-implied positional baseline, so current-season output is mixed with history rather than replacing it. Rates are uplifted for penalty takers, then scaled by expected minutes and by the team's projected goals in that specific fixture.
7 · Defensive & goalkeeping events
Clean-sheet points use the Poisson clean-sheet probability × P(60+). Goals-conceded deductions are integrated over the concession distribution. Keepers get expected saves = save rate × opponent shot volume, so a busy keeper behind a weak defence can still project well.
8 · Bonus model
BPS is projected from a shrunk BPS/90 baseline plus the match-specific contribution (expected goals, assists and clean sheets), then mapped to probabilities of 3, 2 and 1 bonus points.
9 · Distribution, not a single number
Each scoring component is a discrete distribution; they are convolved exactly to produce floor (10th), median, ceiling (90th), blank probability, return probability and 10+ haul probability.
10 · Horizons
Every fixture is projected separately and then aggregated, so blanks and doubles are handled naturally — a double Gameweek simply contains two fixture projections.
Data provenance
- Source
- Official FPL API (fantasy.premierleague.com)
- Data fetched
- Tue, 20 Oct 2026 06:55:02 GMT
- Prediction generated
- Tue, 20 Oct 2026 06:55:08 GMT
- Next deadline
- Fri, 23 Oct 2026 17:30:00 GMT
Statistics are never invented. Where a metric is not published by the data provider it is omitted and flagged rather than estimated silently.
Data — · Model —. FPL projections are estimates, not guarantees. How the model works · Spread betting model · Free bet offers
Frequently asked questions
- How does the FPL expected points model work?
- The model blends market-implied priors with Elo team ratings, runs a Poisson fixture simulation for goals and clean sheets, estimates expected minutes, adds bonus point projections and produces a full points distribution per player.
- Does the model use bookmaker odds?
- Yes. Market prices are used as a prior for team strength and match outcomes, because betting markets are among the best-calibrated public forecasts of football results.
- Why do projections change between updates?
- Projections respond to new information: injuries, suspensions, line-up news, fixture changes and price moves all shift expected minutes and expected points.