How Score Sensei Predicts Football
Model v0.4 (current) · v0.5 beta running in parallel — September 2026
Score Sensei combines historical match analysis with an Elo rating system and Monte Carlo simulation to generate match probabilities and season forecasts. Our approach uses two complementary models — one tuned for near-term match predictions and another optimized for end-of-season projections — each data-driven and continuously validated.
1. Building the Dataset
Everything starts with data. We maintain a historical database of over 60,000 club matches spanning more than a decade, covering the top European leagues and continental competitions. New results are merged in as they happen, and team names are harmonized across sources so that a club's full history is always connected.
For international tournaments like the FIFA World Cup, we use a separate dataset of 8,000+ international matches to build nation-level predictions.
2. Elo Ratings
At the core of our model is an Elo rating system — a method originally developed for chess that has become a standard for measuring team strength in sports. Every team starts at a baseline rating, and after each match their rating is updated based on the result, the margin of victory, and the strength of the opponent.
Stronger opponents are worth more: beating the league leader moves your rating up more than beating a relegation side. We also factor in competition importance — a Champions League match carries more weight than a friendly — and apply a small annual regression to prevent ratings from drifting too far from reality over time.
The result is a single number per team that captures their overall quality, updated chronologically through every match in our database.
3. Rolling Performance Signals
Elo captures long-term quality, but football is also about form, momentum, and context. Alongside Elo, we compute several rolling performance metrics for every team:
- Global averages — a team's weighted scoring and conceding rates across their full recent history, with stronger opponents counting more
- Recent form — scoring patterns from the last handful of matches, capturing hot streaks and slumps
- Home and away splits — some teams are significantly stronger at home; this signal captures that venue effect
Each metric is weighted by opponent quality and competition importance, so scoring three goals against a top-four side counts more than doing the same against a bottom-half team.
4. Dual Prediction Model
We use a dual-model approach because different signals matter at different time horizons:
- Near-term model (next few weeks): emphasizes Elo and recent form — when predicting tomorrow's match, current momentum matters most
- Simulation model (season projections): emphasizes Elo and global averages — for end-of-season forecasts, long-run quality is more predictive than a two-match hot streak
Both models combine their signals into expected goals for each team, then apply a Dixon-Coles correction — a well-known statistical adjustment that improves the modeling of low-scoring outcomes like 0-0 and 1-1 draws. The final output is a probability distribution over all possible scorelines.
5. Monte Carlo Simulation
For league tables and tournament brackets, we go beyond single-match probabilities. We run 10,000 Monte Carlo simulations of the remaining season, randomly sampling goal totals from Poisson distributions for each unplayed match.
Each simulation produces a complete final table. By counting how often each team finishes in each position across all 10,000 simulations, we get probabilities for winning the title, qualifying for continental competition, or being relegated. For knockout tournaments like the Champions League and World Cup, we simulate the full bracket — using actual fixtures and results where available, and modeling unplayed matches from there.
6. International Tournaments
For the FIFA World Cup and other international competitions, we use a dedicated model built on the Dixon-Coles framework — a statistical model that simultaneously estimates attack and defense parameters for every national team using maximum likelihood estimation. This approach naturally handles the challenge of varying opponent quality in international football, where teams play far fewer matches than clubs and face opponents from different confederations.
Tournament simulations follow the official format, including group stages, third-place advancement rules, and bracket constraints.
7. From v27 to v28
From September to early November 2026 the site ran two versions of the club model side by side, with a switch in the top bar. In October the v28 model became the default, and in November 2026 the previous v27 model was retired. v28 is still the default; the switch now offers the v29 beta instead. Both versions were refreshed by the same weekly run from the same match data, so the comparison below measures the club model itself.
v28 came out of a backtest of the model against bookmaker closing odds on 56,000 matches across 13 leagues (2014–2026). Four changes to the Elo engine improved out-of-sample accuracy in every season tested:
- Slower ratings. The Elo K-factor drops from 40 to 18: a single result moves a team's rating less than half as far. The old setting over-reacted to one-off scores.
- Elo only for near-term predictions. The rolling last-five and season-average signals are dropped from the match model; in the backtest they added noise rather than information. They remain in the season-simulation model.
- Promoted teams start lower. A team's first appearance in a second-tier competition starts 200 Elo points below the baseline, so newly promoted or newly arrived clubs are no longer rated as league-average for a season.
- Stronger home advantage. Home advantage in the goal mapping rises from 60 to 100 Elo points, correcting a consistent 4–5 point under-estimate of home wins.
In the backtest these changes closed about a third of the accuracy gap between the model and the betting market. Live results confirmed the backtest: on about 770 matches predicted by both models before kickoff (September–October 2026), v28 scored better on log loss and accuracy, mainly because v27 kept under-estimating home wins, so v28 became the default. We are transparent about what this does not mean: neither version beats the closing odds, and nothing on this site is betting advice.
8. The v29 Beta (November 2026)
v29 is a rebuilt club model, available through the v29 beta switch in the top bar. Before going live it was tested on the 2025–26 season, which it had never seen; all of its settings had been fixed on earlier seasons.
- Attack and defence ratings. Next to Elo, every team carries separate attacking and defensive ratings, updated after each match. This was the largest single improvement.
- Two effects the data supports. Clubs just relegated keep playing above their rating in the second tier (about 10% more goals scored, the “parachute” effect). A team facing an opponent with nothing left to play for scores about 15% more.
- Season tables that look further ahead. For matches months away, v29 leans less on current form and more on a club’s long-run level. Each simulated season also includes some doubt about every team’s true strength.
We also tested momentum, rest days, midweek European games, club-specific home advantage and more; they added nothing beyond the ratings, so they were left out.
On the 2025–26 test season, v29’s next-match forecasts beat v28’s (log loss 1.018 against 1.021). Its season tables were more accurate early in the season, and its relegation odds were better calibrated. Its title odds were no better. Like every version, it remains behind the bookmakers’ closing odds.
Finished matches: in both views, a finished match shows the forecast the model made from what was known before kickoff, so their hit rates and calibration plots can be compared directly. v29 becomes the default only if it beats v28 on live, pre-kickoff predictions.
Competitions We Cover
Score Sensei provides predictions for the following competitions:
- European 1st division: Premier League, La Liga, Ligue 1, Serie A, Bundesliga, Primeira Liga, Eredivisie, Belgian Pro League
- European 2nd division: EFL Championship, 2. Bundesliga, Ligue 2, Serie B, Segunda División
- European cups: UEFA Champions League, UEFA Europa League, UEFA Conference League
- International: FIFA World Cup 2026, CONMEBOL World Cup Qualifying, Copa América, UEFA European Championship
Model Performance
We evaluate our model using calibration analysis: when the model predicts a 30% chance of something happening, does it actually happen about 30% of the time? The calibration plot below shows our predicted probabilities against actual outcomes across thousands of matches. Every point uses the forecast made before kickoff. Points close to the diagonal line indicate well-calibrated predictions.
Leagues Included in the Calibration
- La Liga
- Ligue 1
- Premier League
- Serie A
- Primeira Liga
- Bundesliga
- Eredivisie
- Belgian Pro League
- 2. Bundesliga
- Ligue 2
- EFL Championship
- Serie B
- Segunda División
Changelog
v0.6 beta — November 2026
- v29 beta published alongside v28 (switch in the top bar). It adds attack and defence ratings, the relegated-club and nothing-to-play-for effects, horizon-aware season simulations, and honest pre-kickoff forecasts for finished matches.
- Validated on the 2025–26 season, held out until all choices were fixed.
v0.5 — November 2026
- v27 retired after about 1,000 live pre-kickoff predictions in which v28 scored better.
- Finished matches in the v28 view now show the forecast made before kickoff. They used to be recalculated after the result was known, which made v28 look more accurate than it was: a 57% hit rate shown this season, against 51% before kickoff.
v0.5 — October 2026
- v28 promoted to the default model after outscoring v27 on about 770 live pre-kickoff predictions; v27 remains available through the model switch.
v0.5 beta — September 2026
- Parallel model run: v28 beta published alongside v27 with a site-wide model switch.
- Elo K-factor 40 → 18; near-term match model uses Elo only; second-tier newcomers start 200 points below baseline; home advantage 60 → 100.
- Validated against bookmaker closing odds on 56,000 matches (2014–2026) with a 2015–21 / 2022–26 hold-out.
v0.4 — April 2026
- Introduced Elo rating system for both club and international models, computed from 60,000+ historical matches.
- Implemented dual prediction model: separate near-term and simulation weight configurations, each independently optimized.
- Added Dixon-Coles correction for improved low-score probability modeling.
- Switched to vectorized Monte Carlo simulation (10,000 iterations) for faster league and tournament projections.
- Added FIFA World Cup 2026 predictions with full group stage and knockout bracket simulation.
- International model rebuilt using Dixon-Coles maximum likelihood estimation on 8,000+ matches.
v0.3 — September 2024
- Added more weight to the 5 most recent matches for improved form sensitivity.
- Included Belgian and Dutch leagues.
v0.2 — June 2024
- Fixed home advantage calculation for neutral-ground matches.
- Added team and league strength ratings.
Comments