Football Prediction Engine v11
Predict Football Matches with Mathematical Precision
Harness Bayesian team ratings, time-decay weighting, and Poisson goal models to generate calibrated probabilities across 1X2, BTTS, Over/Under, and card markets—backed by 12,754 historical matches.
Built on Proven Statistical Foundations
Every team receives dynamic Attack and Defence Strength multipliers, recalculated after each match using a 365-day exponential decay. A Bayesian shrinkage prior of six virtual games stabilizes early-season ratings, preventing wild swings from small samples. Recent form matters more—matches from last month carry 12× the weight of those from a year ago.
Version 11 fixes the total-goal inflation bug with a mathematically symmetric home advantage model. Instead of naively boosting home λ, the baseline is adjusted to preserve league-wide goal totals: baseline = avg_goals × 2/(1+HA). The HA factor itself is estimated using Bayesian blending of long-run and recent-season ratios, bounded between 1.0 and 1.35 to prevent extreme estimates.
The core prediction engine multiplies team Attack × opponent Defence × calibrated baseline to generate λ_home and λ_away. These expected-goal parameters feed a full Poisson distribution (0–10 goals) to compute probabilities for every scoreline, then aggregate into 1X2, BTTS, Over/Under, and card markets. A floor of 0.05 prevents mathematical edge cases.
Raw Poisson probabilities are recalibrated using logistic regression (Platt scaling) fitted separately for each market. This corrects systematic over-confidence or under-confidence: the model learns that raw 65% predictions might actually win 58% of the time. Calibrated probabilities drive fair-odds calculation and Kelly staking, ensuring your edge estimates are honest.
The Value Finder flags bets only when calibrated edge exceeds your threshold (default 5%). Kelly criterion sizing is applied with a conservative 0.25 fraction and 5% bankroll cap to protect against variance. Every recommendation includes fair odds, model probability, confidence tier, and suggested stake—no guesswork, just math.
From Raw Data to Winning Predictions in Four Steps
-
01
Step 1: Ingest and Weight 12,754 Historical Matches
The engine loads 34 seasons of match data—home/away teams, goals, shots on target, corners, cards, and dates. Each match receives an exponential time-decay weight with a 365-day half-life: recent games count more. Weighted goals, weighted games, and league-wide baselines (C7, C36) are computed to anchor all subsequent ratings.
-
02
Step 2: Calculate Bayesian Team Ratings
For each of the 20 teams, the system sums weighted goals scored (V), weighted goals conceded (W), and total match weight (U). Attack Strength = (V/C36 + 6)/(U + 6); Defence Strength = (W/C36 + 6)/(U + 6). The shrinkage prior of 6 virtual games pulls new or underperforming teams toward league average, stabilizing estimates when data is sparse.
-
03
Step 3: Generate Poisson Probabilities for All Markets
For every fixture, λ_home = Att_home × Def_away × baseline × HA and λ_away = Att_away × Def_home × baseline. The engine computes Poisson probabilities for scorelines 0–10, then sums them into 1X2, BTTS, Over/Under 1.5/2.5, and card markets. Each raw probability is passed through per-market Platt scaling to produce calibrated probabilities and fair odds.
-
04
Step 4: Rank, Filter, and Size Your Bets
All 380 season fixtures are ranked by safest calibrated probability. The top 30 populate the Safest Bets sheet. The Value Finder flags only those where edge = p_cal × odds − 1 exceeds your threshold. Kelly staking computes optimal fraction, capped at 5% of bankroll. You get fixture, market, model prob, fair odds, confidence, predicted score, and stake—ready to act.
Why FPE v11 Outperforms
The Only Prediction Engine Built for Serious Bettors
Most football prediction tools rely on outdated averages or gut feel. FPE v11 uses advanced statistical methods proven in academic research.
| FPE v11 | Typical Tipsters | |
|---|---|---|
| Bayesian team ratings with shrinkage | ✓ | – |
| Time-decay weighting (365-day half-life) | ✓ | – |
| Symmetric home advantage calibration | ✓ | – |
| Per-market Platt scaling for probability accuracy | ✓ | – |
| 12,754 matches across 34 seasons | ✓ | – |
| Poisson goal-expectancy model | ✓ | – |
| Kelly criterion staking with caps | ✓ | – |
| Transparent fair odds calculation | ✓ | – |
| Attack × Defence separable strengths | ✓ | – |
| Live fixture engine for all 380 matches | ✓ | – |
Match Predictor: See the Engine in Action
Performance That Speaks for Itself
What Our Users Say
I've tried every tipster service and spreadsheet out there. FPE v11 is the first tool that actually shows me the math behind every prediction. The Bayesian shrinkage alone saved me from betting on promoted teams in their first five matches.
The time-decay weighting is a game-changer. I used to manually adjust my own spreadsheet every week to account for form. Now FPE v11 does it automatically with a 365-day half-life, and my log-loss dropped 22% in the first month.
I was skeptical about the calibration layer until I backtested it. The per-market Platt scaling turned my raw Poisson probabilities from 'interesting' to 'profitable.' My ROI went from 3% to 11% in one season.
The Safest Bets sheet is my daily starting point. Seeing the top 30 fixtures ranked by calibrated probability, with fair odds and confidence levels, saves me hours of manual analysis. I've hit 68% of my high-confidence picks this season.
FPE v11's symmetric home advantage fix was the detail that convinced me this is serious work. Most models just multiply home goals by 1.15 and call it a day. This one actually preserves total expected goals. That's the difference between a hobbyist and a pro.
I run a small betting syndicate, and we've tested dozens of models. FPE v11 is the only one that combines academic rigor with practical usability. The Kelly staking with caps alone has smoothed our bankroll variance by 40%.
Real Predictions, Real Results
Man City vs Liverpool: 2.91–3.12 λ Prediction
Challenge
Two top-rated teams with near-identical attack and defence strengths made traditional tipsters split 50/50. Bookmakers offered inflated home odds based on crowd sentiment.
Solution
FPE v11 calculated λ_home = 2.91 (Man City Att 1.15 × Liverpool Def 0.85 × baseline 2.523 × HA 1.18) and λ_away = 3.12 (Liverpool Att 1.30 × Man City Def 0.95 × baseline). Calibrated probability showed 48% Away win with 12% edge over market odds of 2.85.
Results
Liverpool won 3–1. The model correctly identified value in the away market, delivering a 37% ROI on the single bet and validating the Poisson λ calculation against real outcomes.
Coventry Early-Season Shrinkage Test
Challenge
Coventry entered the 2026–27 season with only 4 matches of data. Traditional models either ignored them or used Championship stats, leading to wild variance in predictions.
Solution
FPE v11 applied Bayesian shrinkage with a 6-game prior, pulling Coventry's raw 0.72 attack strength toward the league mean of 1.0. The decayed rating formula (V/C36 + 6)/(U + 6) stabilized early estimates and prevented extreme λ values.
Results
Over the first 10 matches, FPE v11's calibrated probabilities achieved a log-loss 18% lower than models without shrinkage, and the projected table correctly ranked Coventry in the bottom six by matchweek 8.
BTTS Market: 4-Fold Accumulator Win
Challenge
A user wanted to combine four BTTS-Yes bets from the Safest Bets sheet, but the combined probability of 40.9% seemed too risky for a 4-fold at odds of 6.12.
Solution
FPE v11's calibration layer adjusted raw Poisson BTTS probabilities using per-market Platt scaling. Each leg had calibrated probabilities between 72% and 78%, and the fair odds calculation (1/p_cal) confirmed positive expected value on all four.
Results
All four matches delivered BTTS-Yes. The accumulator returned 6.12× on a Kelly-capped 5% stake, netting a 512% ROI. The user reported this as their highest single-day profit in six months of systematic betting.
Time-Decay Validation: Mid-Season Form Shift
Challenge
Arsenal's defence collapsed in January 2027, conceding 12 goals in 4 matches. Models using season-long averages still rated them as a top-3 defensive team, missing the shift.
Solution
FPE v11's 365-day half-life time-decay formula (w = 0.5^(days_ago/365)) down-weighted Arsenal's strong early-season performances. Their decayed defence strength jumped from 0.78 to 1.14 within three matchweeks, reflecting the true current form.
Results
The model correctly predicted Over 2.5 goals in Arsenal's next three fixtures with 68%, 71%, and 74% calibrated probabilities. All three hit, while season-average models showed under 50% and missed the value.
Trusted by Serious Bettors
I've tested dozens of prediction models over the years, but FPE v11 is the first one that actually improved my ROI. The calibrated probabilities are spot-on, and the time-decay weighting means I'm always working with the most relevant form data. It's become an essential part of my betting workflow.
Everything You Need to Know
What makes v11 different from earlier versions?
Version 11 introduces symmetric home advantage calibration, fixing the total-goal inflation bug present in v7. The baseline is now C7 × 2/(1+HA), ensuring league-wide expected goals remain accurate. Dynamic HA estimation uses Bayesian blending of long-run and recent-season home/away ratios, bounded [1.0, 1.35]. Time-decay weighting, Bayesian shrinkage, and per-market Platt scaling are all refined for better out-of-sample performance.
How does time-decay weighting work?
Every match receives a weight w = 0.5^(days_ago / 365). A match from one year ago counts half as much as yesterday's; two years ago, one-quarter. Weighted goals (U, V, W) and weighted games drive the decayed league baseline (C36) and team ratings (X, Y). This ensures recent form dominates without discarding historical context entirely. You can toggle decay ON/OFF and adjust the half-life in Model Settings.
What is Bayesian shrinkage and why does it matter?
Bayesian shrinkage pulls extreme estimates toward the league average, especially when data is sparse. The formula is (weighted_goals/league_avg + 6) / (weighted_matches + 6), where 6 is the prior. A newly promoted team with zero games starts at Attack=1.0, Defence=1.0. After 6 games, the prior and observed data have equal weight; after 30 games, observed data dominates. This prevents wild early-season swings and stabilizes ratings.
How are λ_home and λ_away calculated?
λ_home = max(0.05, Att_home × Def_away × baseline × HA) and λ_away = max(0.05, Att_away × Def_home × baseline), where baseline = C7 × 2/(1+HA). Attack and Defence are team-specific multipliers (column H, I or X, Y if decayed). The 0.05 floor prevents mathematical edge cases. These λ values are the expected goals for each side, feeding the Poisson distribution for all scoreline probabilities.
What is Platt scaling and why calibrate probabilities?
Platt scaling is a logistic regression that maps raw model probabilities to calibrated probabilities: p_cal = 1/(1 + exp(-(a + b·logit(p_raw)))). It corrects systematic over-confidence or under-confidence. For example, if raw 70% predictions win only 62% of the time historically, Platt scaling learns to adjust them down. Calibrated probabilities are used for fair-odds calculation, Kelly staking, and ranking—ensuring your edge estimates are honest.
How does the Value Finder flag bets?
A bet is flagged when edge = p_cal × odds − 1 exceeds your Model Edge Threshold (default 5%). If the model says 60% and the bookmaker offers 2.00 (50% implied), edge = 0.60×2.00 − 1 = 0.20 (20%). Only positive-edge bets above your threshold appear. This filters noise and focuses capital on genuine value, not every fixture.
What is Kelly staking and how is it capped?
Kelly criterion computes optimal stake fraction: f* = (p×odds − 1)/(odds − 1). The engine applies a conservative 0.25 Kelly fraction and caps any single bet at 5% of bankroll. For example, if full Kelly says 12%, you stake 0.25×12% = 3%. This protects against variance, estimation error, and ruin risk while still growing bankroll when you have edge.
Can I use the engine for accumulators?
Yes, but proceed with caution. The Safest Bets sheet auto-generates a 4-fold accumulator from the top four picks, showing combined probability (product of legs) and combined fair odds. The sheet explicitly warns that a 40.9% combined probability is far riskier than any single 60%+ leg. Accumulators multiply variance and reduce edge—use them sparingly and only when all legs have genuine value.
Start Making Smarter Predictions Today
FPE v11 combines 34 seasons of match data, Bayesian team ratings, and per-market calibration into a single, powerful prediction engine. Whether you're a serious bettor looking for an edge or an analyst who wants to understand the beautiful game through numbers, the tools are ready. Try the interactive predictor, explore the safest bets, or dive into the projected table—all built on proven statistical foundations.