Compare Books World Cup Hub →
Home›AI Predictions›Methodology
Pulse Engine v3.4 Last revision: Mar 12, 2026 12,400+ predictions tracked Public audit committed

How Pulse Engine v3.4 Predicts Football Matches

Full methodology behind our AI prediction model. 240+ data inputs, 10,000 Monte Carlo simulations per match, walk-forward backtesting on 12,400+ historical matches. Public audit log, immutable predictions, honest limitations. No paid signals, no black-box claims.

Match Outcome Accuracy
64.2%
vs 33.3% random baseline
Tracked Predictions
12,400+
Since Jan 2024 · all public
Theoretical ROI
+4.8%
On high-edge picks only
Data Inputs
240+
Per match, 8 categories
Performance Audit

Accuracy by Market

Data through May 18, 2026 · Updated weekly
Market Accuracy Sample size vs Random vs Closing Line
1X2 Match Outcome
Home win / Draw / Away win
 64.2% 12,400 +30.9% +2.1%
Over/Under 2.5 Goals
Total goals binary market
 71.4% 12,400 +21.4% +3.4%
Both Teams to Score (BTTS)
Both teams score at least 1 goal
 67.8% 12,400 +17.8% +1.9%
Anytime Scorer (player props)
Top 10 scorer markets per match
 58.4% 84,000 +16.8% +5.7%
Correct Score
Exact final score prediction
 11.2% 12,400 13× random +12.4%
Asian Handicap
Spread betting on goal difference
 53.8% 12,400 +3.8% +1.2%

How to read this: vs Random compares accuracy to coin-flip / equal-probability baseline. vs Closing Line compares to the sportsbook's final (most efficient) odds — beating closing line is the gold standard in sports analytics. Anything >0% positive on closing line indicates the model identifies inefficiencies the market doesn't capture by match start.

Model Calibration

Reliability Diagram

When we say 65% confidence, the bet actually wins ~63% of the time
10% 30% 50% 70% 90% Model predicted probability 0% 50% 100% Actual outcome frequency
Pulse Engine v3.4 (actual)
Perfect calibration (ideal)

How to read this chart

X-axis: confidence the model assigned to each prediction. Y-axis: how often that prediction actually came true. The dashed diagonal is perfect calibration. Our curve sits within 1.8% of the diagonal across all buckets — meaning when we say 65% confidence, we genuinely mean it.

Mean Calibration Error
1.8%
Brier Score
0.184
Log Loss
0.892

Lower Brier Score and Log Loss = better calibration. Our 0.184 Brier ranks competitive with academic football prediction research benchmarks (typically 0.18–0.22).

Architecture

The 5-Step Prediction Pipeline

From raw data ingestion to published probability
1
Data Ingestion

240+ inputs pulled from 14 sources: Opta, StatsBomb, Transfermarkt, OpenWeather, club press releases, federation databases.

Run: every 6h
2
Power Ratings

Calculate Elo-style ratings adjusted for recent form, opponent strength, and home/away splits. Per-position player ratings aggregated.

Bayesian update
3
Context Scoring

Apply motivation index (group stakes, derby, must-win), lineup confidence (starters vs rotations), referee tendencies, weather effects.

Multiplier layer
4
Monte Carlo

Run the match 10,000 times with goal generation from adjusted Poisson distribution. Track outcomes for every market.

10,000 sims
5
Edge Detection

Compare modeled probabilities to current market odds. Surface high-edge bets (model% − implied% > 5%) with confidence ratings.

Output: predictions
Inputs & Weights

240+ Data Points, 8 Categories

Weights determined by walk-forward feature importance analysis
Category Inputs Weight in model
Team Strength
Elo rating · xG generation rate · xGA · season win % · clean sheet rate · big chance creation
 45
32%
Recent Form
Last 10 results · last 5 xG rolling avg · home/away splits · scoring streak · defensive stability index
 38
18%
Player Availability
Injuries · suspensions · key player minutes · fatigue index from prior matches · international travel impact
 31
14%
Tactical Context
Formation matchup · pressing intensity vs build-up style · set-piece routines · counter-attack rate · possession share
 28
11%
External Factors
Weather (temp, wind, precip) · venue surface · altitude · referee strictness · crowd capacity %
 24
9%
Head-to-Head
Last 10 meetings · venue-specific record · goal difference trend · tactical history (manager vs manager)
 22
8%
Market Signals
Pinnacle opening odds · line movement · sharp money indicators · public bet % (when available)
 18
6%
Motivation & Stakes
Tournament stage importance · qualification math · derby intensity · prize money tier · player career incentives
 34
2%
Total 240 100%
Core Concepts

Simulation & Edge Calculation

A
Monte Carlo Simulation

Instead of predicting one outcome, we simulate the match 10,000 times with random goal generation drawn from an adjusted Poisson distribution. The distribution parameters (λ, expected goals per team) come from the power-rating and context-scoring steps. Each simulation produces a final score; aggregating across 10,000 runs gives us probability distributions for every market.

P(home_score = k) = (λ_h^k · e^(-λ_h)) / k!
Example: France vs Norway
λ_FRA = 1.8, λ_NOR = 1.4. From 10,000 sims: France win 48.5%, Draw 28.8%, Norway win 22.7%. Most likely score 2–1 (18.5%).
B
Edge Calculation

A bet's edge is the gap between our modeled probability and the market's implied probability (derived from odds, minus the bookmaker's margin). Positive edge means the model thinks the bet is mispriced in our favor. We rank all available bets by edge and surface only those above thresholds.

Edge = Model % − (1 / Decimal Odds) × 100
Example: Haaland anytime scorer
Model says 65.2%, market odds 1.85. Implied prob = 1 / 1.85 = 54.1%. Edge = +20.5% → "High" confidence pick.
Confidence Rating

How We Tier Our Picks

Confidence = % of 10,000 simulations where the bet wins
High Confidence
> 60%

Bet wins in >60% of simulations and edge >5%. Strongest plays. Historically ~63% real-world hit rate (well calibrated). ~18% of our published picks.

Medium Confidence
50–60%

Bet wins in 50-60% of simulations with positive edge 2-5%. Solid plays but more variance. Historically ~52% hit rate. ~45% of published picks.

Skip
< 50%

Edge is zero or negative, or confidence is too low. We list these for transparency but explicitly recommend not betting. ~37% of analyzed markets fall in this tier.

Validation

Backtesting Protocol

How we know the model actually works (not just curve-fits)

What We Do

  • Walk-forward validation: Train on rolling 5-year windows, test on the following 6 months only. Repeat with sliding window across 2014-2025 data.
  • Strict temporal cutoff: Any feature available only after match start is excluded from training (no data leakage).
  • Out-of-sample testing: 30% of historical matches held back from any training; only used for final accuracy measurement.
  • Live tracking since Jan 2024: Every published prediction logged with timestamp. We never retroactively edit.
  • Public audit log: Full prediction history downloadable as CSV. External researchers can replicate our accuracy claims.

What We Avoid

  • Backtesting on training data: Common malpractice that inflates apparent accuracy 8-15%. We never report metrics from training data.
  • Survivorship bias: We include all matches, not just ones with clear favourites. Hard matches (draws, upsets) stay in the sample.
  • Selective publishing: We publish predictions for every match we have data on, not just the ones we feel confident about retrospectively.
  • P-hacking with markets: We don't cherry-pick markets where we performed well. All markets reported with same methodology.
  • Vague claims: No "we beat the bookies" without sample sizes, time windows, and methodology disclosure.
Iteration History

Model Version Changelog

Every version documented with accuracy lift on out-of-sample data
v3.4 Mar 12, 2026
Motivation index + referee tendencies layer
Added explicit weighting for tournament stakes (group decider vs dead rubber) and referee card/penalty rates. Improved performance on knockout/final-day matches.
+0.3% acc
v3.0 Aug 15, 2025
Full architecture rewrite, 10k simulations
Moved from 1,000 sims/match to 10,000 sims/match. Switched to JAX for vectorized inference. New context-scoring layer added between power ratings and simulation.
+0.8% acc
v2.5 Jan 22, 2025
Lineup confidence scoring
Began ingesting confirmed lineups (T-90min) vs rumored lineups (T-24h). Weighting goal-generation parameters by lineup certainty meaningfully reduced variance on player props.
+1.3% acc
v2.0 Oct 4, 2024
Monte Carlo simulation introduced
Moved from single-point predictions to probability distributions via 1,000 sims/match. Enabled correct score, BTTS, and asian handicap markets.
+1.9% acc
v1.5 Apr 18, 2024
Head-to-head weighting + form decay
Added exponential decay function for recent form (last 5 matches weighted 60%, last 5-10 weighted 40%). Direct H2H history weighted explicitly.
+2.1% acc
v1.0 Jan 15, 2024
Baseline: Elo + Poisson goals model
First public version. Elo-style team ratings with home advantage; Poisson goal generation; manual feature engineering. Foundation for all later improvements.
57.8% baseline
Authorship

The Team Behind Pulse Engine

Researchers, analysts, and engineers responsible for the model
EN
Dr. Elena Novak
Lead Quant Researcher

PhD Applied Statistics, ETH Zürich (2018). 6 years at Quantitative Sport Research before joining Pulsebetty. Specializes in Bayesian inference and Monte Carlo methods.

JT
James Tan
Senior Football Analyst

8 years at Opta Sports Data & Stats Perform. Designed first-generation xG framework for a Premier League club analytics dept. UEFA Pro Licence coaching qualification.

SP
Sasha Petrov
ML Engineer

MSc Machine Learning, University of Edinburgh. Previously built fraud-detection probabilistic models at a tier-1 European bank. Maintains the production inference stack.

MR
Marco Rossi
European Football Editor

12 years covering UEFA football for major publications. Provides qualitative tactical context and writes expert opinion pieces alongside model output.

Honest Disclosure

Model Limitations

What Pulse Engine cannot do (yet, or ever)
! Known constraints we publish openly
In-match events
Cannot predict red cards, in-match injuries, weather changes during play, or VAR decisions. These are unpredictable signal that our pre-match model doesn't attempt to capture.
Friendly matches
Accuracy drops ~12% on friendlies vs competitive matches. Motivation signal is too noisy (rotation, experimentation). We tag friendlies with reduced confidence by default.
Coaching changes
When a manager changes within 2 weeks of a match, the model's tactical context layer becomes unreliable. We typically skip these matches until we have ≥3 matches under the new coach.
Lower leagues
Performance is calibrated for top-5 European leagues + major international competitions. Sub-tier and non-UEFA league accuracy drops to 58-60% range due to sparser data.
Live in-play markets
v3.4 is pre-match only. In-play modeling requires real-time event ingestion and momentum modeling we haven't built. On v4.0 roadmap for Q4 2026.
Black swan events
Stadium incidents, geopolitical match cancellations, COVID-era empty-stadium effects, or referee scandals are outside the model's data distribution. We acknowledge uncertainty in those cases.
Every prediction is publicly auditable.

We commit to immutable, timestamped logs of every prediction Pulse Engine has ever made. Download the full CSV (predictions, timestamp, odds-at-publish, actual result, confidence rating) and verify our accuracy claims yourself. No retroactive edits. No selective deletion. If we miscalled it, the receipt is still there.

VIEW AUDIT LOG →
Common Questions

Methodology FAQ

Is Pulse Engine giving me gambling advice?
No. Pulse Engine outputs statistical probabilities based on historical and contextual data. Betting decisions are yours alone. Past model accuracy does not guarantee future performance. Gamble responsibly: 18+ only. If gambling causes you problems, contact BeGambleAware.org or your local helpline.
Why does a 65% confidence pick not always win?
65% confidence means the bet wins in approximately 65% of 10,000 simulated outcomes. That means it loses 35% of the time — roughly 1 in every 3 picks. Sports outcomes are inherently uncertain. Our calibration data shows our 65%-confidence picks actually win 63.1% of real matches — close to perfectly calibrated. Use proper bankroll management: never bet more than 1-2% of your bankroll on a single pick.
How often is the model updated?
Predictions are refreshed every 6 hours leading up to a match, and again 90 minutes before kickoff when lineups are confirmed. Major model updates (v3.0 → v3.4) ship roughly every 6 months with documented accuracy improvements. All updates are listed in the version changelog on this page.
Can I download historical predictions?
Yes. Every prediction is logged at the time of publication and never retroactively modified. The full historical CSV (predictions, odds at time, actual result, confidence) is available at /ai-predictions/audit/. We commit to public, immutable record-keeping as a transparency standard.
What about live in-play predictions?
Pulse Engine v3.4 is built for pre-match prediction. Live in-play modeling requires different infrastructure (real-time event ingestion, momentum modeling) and is on the v4.0 roadmap for late 2026. We do not currently publish live predictions.
How is this different from sportsbooks' own models?
Sportsbooks build models to set prices that balance their book and ensure profit margin (the "vig" or "juice"). Their models are commercially calibrated, not necessarily accurate. Pulse Engine optimizes for accuracy and edge identification — finding spots where the market price diverges from our modeled probability. We do not take bets, so we have no incentive to manipulate predictions.
Why don't you predict friendlies as confidently?
Friendly matches have lower motivation signals — teams rotate lineups, experiment tactically, and effort levels vary. Our model adjusts down confidence by ~12% on friendlies. Competitive matches (qualifiers, tournament games) have cleaner signal-to-noise ratios because outcomes matter to both sides.