Our Methodology
Transparent AI analysis. Real verified results. No cherry-picking.
Hybrid: statistics + AI
Every metric has a source. Every formula is documented. Every outcome is verified automatically. The goal is not to pick winners blindly — it's to understand why and how.
Data-first
We start from raw facts: fixture data, xG, odds, squad news. No assumptions before the data says so.
Quantitative core
Poisson model for probability, expected value calculation, Kelly criterion for stakes. Math runs before the AI opinion.
AI synthesis
AI contextualises numbers — current form, squad quality, venue dynamics — and flags conflicts between math and context.
Everything is verifiable. See the full track record →
The 5-Step Prediction Pipeline
Each prediction goes through the same deterministic sequence. No shortcuts, no skipped steps.
Data Collection
No AI · facts onlyDeterministic data ingestion — no AI involved at this stage. Every number is sourced from external APIs.
- API-Football: fixtures, team stats, H2H, current form, injuries, lineups (700+ leagues)
- Understat xG: expected goals — daily update + our own historical xG base
- Live odds from 6 bookmakers: Melbet, Winline, Pari, Betwinner, Starzbet + aggregated odds feed
Analysis Generation
AI — expert layerAI generates qualitative match analysis from the collected data: attack vs defence assessment, current form dynamics, H2H on specific venue, squad optimality.
- Attack vs. defence quality comparison for both teams
- Current form dynamics (last 5 matches weighted by recency)
- Head-to-head record on this specific venue
- Squad optimality — key player availability factor
Quantitative + AI Betting Analysis
Math + AITwo parallel tracks that are then compared for consistency.
- 3A — Math: fair outcome probabilities via Poisson model, expected value per market, Kelly stake with 0.25 safety multiplier
- 3B — AI: synthesises all prior context and 3A results; can override or confirm the math
- Conflict report: if 3A and 3B diverge significantly, confidence is automatically reduced
Final Review
AI — quality gateConsistency check across all prior steps. AI assigns the final verdict and confidence level, catching contradictions or anomalies.
- Cross-checks Steps 2 and 3A/3B for contradictions
- Assigns final confidence score (clamped 5–95%)
- Sets recommendation strength (low / medium / high)
Publication
Structured outputHuman-readable analysis rendered on the site plus structured JSON consumed by the frontend for live odds overlay, confidence ring, and auto-verification.
- Natural-language analysis visible on prediction page
- Structured JSON: market, prediction value, confidence, Kelly stake
- Auto-verification triggers when match result is available
How confidence is calculated
A single number — clamped between 5% and 95% — represents how reliable we think a prediction is.
Formula
Confidence = 0.4 × Data Quality + 0.3 × Market Consensus + 0.2 × Historical Stability + 0.1 × Context
Result clamped to [5%, 95%]
Data Quality
Sample size, freshness, xG source quality
Market Consensus
Gap between our xG-Poisson probabilities and market odds
Historical Stability
Coefficient of variation of historical xG — lower CV = higher score
Context
Multiplicative product of situational penalty factors
Weight distribution
Live calibration proof
Higher confidence really does win more often. Nothing is hidden.
No calibration data available yet.
When confidence decreases
Non-standard conditions reduce the confidence score multiplicatively. The formula: penalty = venue × motivation × squad.
Venue
Motivation
Squad
Example calculation
A relegation battle (−10%) on a neutral ground (−5%) with 2 key players out (−10%): 0.90 × 0.95 × 0.90 = 0.769 — the context component is multiplied into the final confidence formula, reducing the overall score by roughly 23%.
Profit comes from value, not win rate
We only recommend a bet when our probability estimate is higher than what the bookmaker odds imply — by at least 5%.
Value Formula
Value% = (our probability × bookmaker odds − 1) × 100
Minimum threshold: 5% value to recommend a bet
Kelly Criterion
Optimal stake fraction = (edge / odds − 1) / (odds − 1). We apply a 0.25 safety multiplier to reduce variance for conservative bankroll growth.
Markets scanned
Updated on a rolling cycle — results verified each time a match concludes.
Example: 1X2 market
Our model assigns 32% probability to an away win. Bookmaker offers odds of 4.20 (implied probability: 23.8%).
Value = (0.32 × 4.20 − 1) × 100 = +34.4% — well above the 5% threshold.
This is why our 1X2 win rate sits around 30–31% yet the ROI is positive: at avg odds ~4.0, even a 30% win rate produces profit. Value betting is about expected value, not about winning more than you lose.
How We Find Value Bets
Transparent, math-driven, no cherry-picking. Every value bet passes the same five-step filter — here's exactly what that means.
Model probabilities
Every match goes through our 5-step AI analysis pipeline. The quantitative core derives fair probabilities for 1X2, Over/Under 2.5 and BTTS from xG data, team form, head-to-head patterns and squad availability.
Odds scan
We scan current bookmaker odds for each analysed market and take the best available price. Excluded bookmakers are filtered per region and licensing requirements so only legitimate, accessible prices are considered.
Value calculation
A bet has value when our model probability exceeds what the odds imply. Three numbers tell the full story:
Formula
Implied probability = 1 / odds
Value % = (model probability − implied probability) / implied probability × 100
EV = model probability × odds − 1
Strict filters (honesty gate)
Every opportunity passes a hard filter before appearing on the page. We drop anything that doesn't meet all criteria — no cherry-picking:
- Minimum value ≥ 5%
- Odds between 1.50 and 15.0
- Confidence score ≥ 0.7
- Opportunities above 65% value are discarded as likely data errors
- Every opportunity expires at kickoff
Stake sizing
Kelly Criterion with a conservative 0.25 multiplier, capped at 5% of bankroll. The formula targets long-term bankroll growth while limiting variance:
Formula
Kelly fraction = ((odds − 1) × p − (1 − p)) / (odds − 1)
stake = Kelly × 0.25, max 5% of bankroll
What you see on the odds page
Probability ring
Model probability of the outcome — a math-derived number, not a hype score or editorial rating.
Value %
Our edge over the bookmaker's implied probability. Positive means the price is better than the true odds.
Risk & stake
Risk rating (low / medium / high) and a conservative Kelly stake suggestion as a percentage of bankroll.
Value betting is long-term and probabilistic
Mathematically sound value bets still lose. The edge plays out over hundreds of bets, not individual outcomes. No guarantee of profit per prediction. Bet responsibly — treat every stake suggestion as a reference, not a mandate. 18+.
See live value betsHow the Track Record is calculated
Not every prediction we publish is counted as a staked bet. Here is exactly which picks qualify, and the formula used to turn them into ROI.
What counts as an "Official Tip"
The headline ROI and win rate above — and everywhere else on this site — only count picks that pass all four of these conditions:
Primary pick
The main recommendation for the match — not an alternative market or a side note.
High confidence
Confidence level must be classified High, not Medium or Low.
Model probability ≥ 50%
We only stake on outcomes our model believes are more likely than not.
Positive expected value (EV ≥ 0)
The price must offer at least breakeven expected return given our probability.
This deliberately excludes two kinds of picks that would otherwise distort the numbers: no-value "chalk" (short-odds favourites with no edge) and sub-50% longshots (low-probability outcomes that look exciting but lose more often than they win). Both are still published as analysis on the site — readers can see the full reasoning — but they are not counted as staked bets in the official ROI.
ROI Formula
Stake = 1 unit, flat, on every official tip — no progressive or variable staking.
Win profit = (decimal odds − 1) × 1 unit
Loss = −1 unit
Push (void/cancelled) = 0, excluded from win rate
ROI = (Σ profit / Σ staked) × 100
Win rate = wins / (wins + losses) × 100
Flat 1-unit staking — not the Kelly stake sizing used for individual bet recommendations — is used here so the headline ROI reflects pure picking accuracy, not stake-sizing skill.
We show the losing markets too
Per-market performance is published openly, including markets that are currently unprofitable — we do not hide or exclude a market just because its ROI is negative. The same qualifying rules (primary, High confidence, prob ≥ 50%, EV ≥ 0) apply identically across every market; nothing is curated after the fact to flatter the headline number.
Verified Track Record
Live per-market breakdown from the same data powering our home page — official tips only, as defined above. No cherry-picking; every completed bet is counted.
Honest about what we are and aren't
Automated verification
Results are verified automatically — every hour we check for completed matches and update win/loss/push status. The full daily cycle runs once per day. No manual selection.
No guarantees
Football is fundamentally unpredictable. Even mathematically sound value bets lose. We provide a transparent edge over time — not a guarantee per prediction.
System evolving
Models improve as we collect more data. Past ROI does not guarantee future ROI. We document methodology changes and do not rewrite history.
Bet responsibly
Treat each tip as one input, not a copy-paste decision. Never bet more than you can afford to lose. Use our Kelly stake only as a reference, not a mandate.
Frequently asked questions
Now You Know How It Works — See the Results.
Browse today's AI-powered predictions built on the exact methodology above. Every tip backed by a transparent, verified track record.