Methodology

Seer9 uses a quantitative probability model — not AI predictions or expert picks — to estimate the likelihood of various World Cup match outcomes. It then compares these estimates against bookmaker market odds to surface possible pricing differences.

We do not predict the future. We estimate probabilities and show where market prices may differ from model probabilities.

Each team has an Elo rating — a numerical strength estimate computed from real match history (World Cup, Euros, qualifiers) using the standard Elo formula. K-factors vary by competition importance: K=60 for World Cup finals, K=50 for continental championships, K=30 for qualifiers, K=20 for friendlies.

For example, a team rated 2100 playing against a team rated 1700 has a significant strength advantage, which the model reflects in higher expected goals.

Data source: Football-Data.org match history (WC 2022, WC 2026, Euro 2024).

The model adjusts expected goals based on each team's last 5 completed matches, derived from real results. Recent wins increase the expected goal rate slightly; recent losses decrease it.

More recent matches are weighted more heavily using exponential decay. Both results (W/D/L) and goal difference contribute to the form adjustment.

Data source: Football-Data.org completed match results.

Key player absences (injuries, suspensions) adjust team strength. 56 key players across 20+ teams have impact ratings (0.0–1.0). Missing attackers reduce xG by up to 25%; missing defenders increase opponent xG by up to 20%. Multiple absences stack with diminishing returns (each subsequent at 70% of previous). Elo is also adjusted — up to 80 points per missing player.

Data source: API-Football (api-sports.io) real-time injury and suspension data.

Using the Elo-based strength multiplier and form adjustment, the model calculates an expected goals value (lambda) for each team. This represents how many goals a team is expected to score on average in this specific match.

The baseline is approximately 1.35 goals per team per match (World Cup average). For knockout matches, a small reduction factor is applied to reflect typically tighter, more defensive play.

The expected goals for each team are fed into a Poisson distribution — a standard statistical model for counting events (goals) that occur at a known average rate.

The model generates a 10×10 matrix of score probabilities: the likelihood of every scoreline from 0-0 through 9-9. Each cell represents the probability that Team A scores exactly X goals and Team B scores exactly Y goals.

Standard Poisson assumes goals are independent between teams, which underestimates low-scoring draws. The Dixon-Coles (1997) correction applies a tau factor to 0-0, 1-0, 0-1, and 1-1 scorelines using a rho parameter (typically around -0.13). Rho is context-adaptive: more negative for knockout matches (more cautious play) and less negative for high-scoring contexts.

From the corrected score distribution matrix, the model derives probabilities for market types:

• Match winner: Sum all scorelines where A wins, draws, or B wins. Compared against real bookmaker odds.
• Over/Under 2.5 goals: Sum scorelines with 3+ total goals vs 0-2 total. Compared against real bookmaker odds.

In live mode, only markets with real bookmaker odds are shown. Data source: The Odds API (consensus across multiple bookmakers, vig removed via normalization).

For knockout matches, if 90 minutes ends in a draw, extra time is modeled as a reduced-intensity Poisson process (30 minutes at 60% intensity). If still drawn, penalties are simulated via Monte Carlo (10,000 trials) with a 73% base conversion rate adjusted by Elo difference. The result: full advancement probabilities including the probability of extra time and penalties.

The model compares its estimated probability against the market-implied probability for each outcome:

edge = model_probability - market_implied_probability

A positive edge means the model estimates a higher probability than the market price implies. A negative edge means the market prices the outcome higher than the model suggests.

The model assigns a confidence level based on:

• Elo difference magnitude (larger gap = more predictable)
• Whether recent form data is available
• Overall data quality
• Size of the edge itself

Signal labels are assigned as follows:

• Potential edge:Edge > 5 percentage points and confidence is at least medium.
• Fairly priced: Edge within ±5 percentage points of market.
• Avoid:Edge < -5 percentage points (market prices higher than model estimate).

These labels are descriptive, not prescriptive. "Potential edge" does not mean "buy". "Avoid" does not mean "sell". They describe the relationship between model estimates and market prices.

All predictions are recorded. After 50+ resolved predictions, the system computes Brier scores and a 10-bucket calibration curve. If systematic over- or under-confidence is detected, future probabilities are adjusted accordingly. This feedback loop improves accuracy over time.

For in-progress matches, remaining goals are modeled as Poisson with adjusted intensity based on time remaining. Red cards reduce the carded team's scoring rate by 30% and boost the opponent by 20%. Leading teams late in the game play more defensively (reduced lambda). Impossible scores (going backwards from current score) are zeroed out.

• Does not account for tactical formations or style matchups
• Does not use real-time xG data from match tracking
• Does not account for weather, travel, or referee tendencies
• Does not place trades or recommend trade sizes