Methodology — How Our Predictions Work

1. System Overview

System name: Football Predictions Engine v3
Operator: Alexey Andrianov (independent publisher)
Launched: 2024
Operating location: Kyiv, Ukraine
License: Proprietary — © Football Predictions AI

2. Intended Use

Football Predictions AI is an informational and entertainment publication for football fans aged 18 and over. The predictions presented are probability estimates derived from public bookmaker market data — they are NOT guarantees of match outcomes.

Appropriate uses

Tracking match probabilities and form before kickoff
Comparing our published picks against your own analysis
Following daily football statistics and historical performance
Educational reference for understanding probability vs. odds

Out-of-scope uses

Real-money decisions — do not use our picks as the sole basis for any betting decision. Independent verification is essential.
Professional gambling — our model reflects bookmaker consensus; it does not provide an edge over the market.
Financial planning — predictions are entertainment, not investment advice.

3. Data Sources

All match data — fixtures, line-ups, statistics, and bookmaker odds — is sourced from a single industry-standard provider:

Provider: API-Football
Coverage: 180+ leagues, 70+ countries
Features per fixture: ~50–100 numeric inputs — team form, head-to-head, home/away splits, expected goals (xG), squad availability, league standings, momentum indicators
Bookmaker odds: 15+ major operators per market, snapshot pre-kickoff
Refresh cadence: Continuous — odds re-fetched as bookmakers publish updates, predictions recalculated accordingly

We do not use third-party prediction APIs, "insider" sources, or hand-crafted scouting data. Everything you see on this site can be reproduced from API-Football data + the rules described below.

4. Prediction Engine — How Picks Are Computed

Important: we do not train our own machine-learning models. Our prediction engine is a deterministic statistical aggregation of publicly available bookmaker odds. This section describes exactly how it works.

Step 1 — Implied probability from odds

For each market (1X2, Over/Under, BTTS, Double Chance, Asian Handicap, Half Time, Half Time / Full Time, Correct Score, Corners, Cards, Goalscorer), we collect odds from 15+ bookmakers via API-Football and average them per outcome.

Implied probability per outcome: P(outcome) = 1 / average_odd

Step 2 — De-vigging (margin removal)

Bookmaker odds include a built-in margin (the "vig") that makes implied probabilities sum to more than 100%. We normalize so the probabilities of mutually exclusive outcomes sum to exactly 100%:

P_fair(outcome) = P(outcome) / Σ P(all outcomes)

Step 3 — Pick selection rules

For each market we apply deterministic selection rules. Examples:

1X2: pick the outcome with the highest de-vigged probability; special case: if the home–away split is <5% and draw >28%, the model picks the draw.
Over/Under: pick the line/direction with the highest probability after de-vigging.
Asian Handicap: pick the line aligned with the 1X2 favourite whose odds are closest to 2.0 (balanced risk).
Half Time: pick the HT outcome aligned with the 1X2 favourite, applying de-vigged probabilities for HT-only odds.
BTTS, Double Chance, Correct Score, etc.: analogous deterministic rules — full per-market logic visible in the source code.

Step 4 — Confidence score

Confidence is the de-vigged probability rounded to an integer percent. A pick with 72% confidence means the bookmaker market consensus assigns that outcome a 72% chance after margin removal.

Step 5 — "Bet of the Day" selection

From all available picks across upcoming fixtures, the system surfaces the top 1–5 by confidence as "Bet of the Day" and "Hot Picks". Selection is fully automated; editorial review is described in section 6.

5. Article Generation — How Match Reports Are Written

Article texts on this site are drafted by pre-trained Large Language Models. We do not train, fine-tune, or operate our own LLMs.

Primary model: Qwen3-30B-A3B (self-hosted via vLLM)
Fallback model: Azure OpenAI GPT-4o
LLM inputs: The pre-computed prediction + match context (teams, league, form, head-to-head, recent results) + structured statistics from API-Football
LLM outputs: Narrative description of the prediction + statistical context (HTML article body)

Critical distinction: predictions (the picks, confidence numbers, recommended lines) are NOT generated by the LLM. They come from the statistical engine described in section 4. The LLM produces only the narrative wrapping around the prediction.

6. Editorial Review

Editorial responsibility is held by Alexey Andrianov personally. The review process:

Daily — before publication

Bet of the Day pick reviewed against pre-match news (injuries, suspensions, lineup changes that surface in the final hours)
Anomalous picks (e.g. suspiciously good odds on obscure leagues) are flagged for re-evaluation or held back
Confidence levels cross-checked for sanity (an 95%-confidence pick on a mid-table match warrants a second look)

Weekly

Random spot-check of 2–3 LLM-generated articles for factual accuracy (team names, player names, league context, historical references)
Screening for hallucinations — LLMs occasionally fabricate players or statistics; flagged articles are regenerated
Loss analysis — unexpected losses examined for systematic gaps in the model (e.g. recurring failures on a specific league or bet type)

Monthly

Performance audit — per-league and per-bet-type accuracy breakdown reviewed against historical baseline
Methodology updates — selection thresholds and rules tuned where the data justifies it; changes recorded in the update log (section 10)

All published predictions carry a timestamp recorded before kickoff, making editorial decisions verifiable retrospectively.

7. Performance & Accuracy

Performance is tracked transparently on our public stats page — wins, losses, pushes, and per-bet-type breakdowns are updated daily with no cherry-picking.

Overall "Our Pick" win rate: 60.3% (across 16,179 tracked predictions)
Tracking period: Since launch (2024) — full historical record retained
Confidence calibration: Picks at the 70% confidence level historically win approximately 68–72% of the time; calibration is monitored monthly

Live performance dashboard: /stats

8. Known Limitations

Honesty about what the model cannot do is as important as describing what it does. Our predictions will not be optimal in the following circumstances:

Late-breaking injuries or lineup changes — we capture an odds snapshot a few hours before kickoff; if a key player is ruled out 30 minutes before the match, our prediction may not reflect it.
Match-fixing patterns — we have no mechanism to detect fixed matches. Anomalous bookmaker behaviour would not be flagged.
Weather and pitch conditions — not consistently available across all 180+ leagues, so not used as a model input.
Referee assignments — we do not factor in individual referee tendencies (high cards / penalties / time added).
Player motivation & dressing-room state — public information only; we do not have insider knowledge of squad morale, contract disputes, etc.
Tail risk — a 75%-confidence pick still loses 25% of the time. Past performance does not guarantee future results.
Lower-league accuracy — bookmaker odds in obscure leagues are less efficient; our accuracy is meaningfully lower in 3rd-tier and women's competitions than in top European leagues.
Market = us — because we aggregate the bookmaker market, we do not "beat" it; we reflect it in a friendlier format.

9. Bias, Fairness & Conflicts of Interest

No bookmaker partnerships — we do not receive commissions, affiliate fees, or sponsored placements from any bookmaker. Our recommendations are not paid placements.
No cherry-picking — every published pick is recorded with a timestamp before kickoff and the result tracked. Losing picks are not hidden.
No pay-to-promote — teams, leagues, or fixtures cannot pay to be featured.
Coverage bias — we cover more top-tier leagues than lower divisions because that's where bookmaker coverage is richest. This is a structural bias, not editorial choice.

10. Update Log

v3 — 2026-04: Added corner and card markets, refined Asian Handicap alignment to 1X2 favourite
v2 — 2025-12: Switched primary LLM to Qwen3-30B-A3B (self-hosted via vLLM); GPT-4o retained as fallback
v1 — 2024-06: Initial public release — 1X2, Over/Under, BTTS, Double Chance, Half Time, HTFT, Correct Score

11. Responsible Use

Betting carries real financial and psychological risk. If you choose to bet:

🔞 You must be 18 or over (21+ in some jurisdictions). Confirm local age limits before betting.
Set a strict budget and never bet money you cannot afford to lose.
Treat predictions as probability estimates, not certainties. A 75% confident pick still loses 1 in 4 times.
If you feel betting is becoming a problem, stop and reach out for help.

Free, confidential help is available: Responsible Gambling resources

12. Verifiability & Contact

Everything described above can be verified independently:

Every prediction is timestamped before kickoff and stored permanently.
Live accuracy at /stats — updated daily, win/loss/push records visible per bet type.
API-Football is a verifiable third-party data provider with public documentation.
Operator identity verifiable on X and LinkedIn.

We welcome research collaboration and methodology questions: [email protected]