Backtheo-multifactor-2026.2 · 1,000 simulations

Theo AI · Prediction model

How Theo thinks.

Every prediction is a transparent chain: real-world signals become weighted, sourced factors; those set expected goals; a Poisson model produces every scoreline and market; and a Monte Carlo simulation plays the tournament out thousands of times. Nothing is a black box — open any match to see the full math.

Germany

20.3% title odds

The pipeline

From data to prediction

1
Aggregate signals
Form, injuries, player stats, venue, weather/heat, rest, and market odds — each mapped onto canonical teams.
2
Score every factor
Each signal becomes an explainable contribution in rating points, toward one side or the other.
3
Compose the edge → xG
Factors sum to a net rating delta that sets each side's expected goals.
4
Poisson scoreline
A bivariate-Poisson grid yields win/draw/win, every scoreline, BTTS, over/under, and projected scorers — all consistent.
5
Monte Carlo bracket
Thousands of simulations sample group results and knockout winners for advancement and champion odds.
6
Explain it
Every number is reproducible and narratable — ask Theo why on any match to reveal the components.

The formulas

The math, in the open

Edge → expected goals

xG = base · e^(Δ / 600) · form

The net rating edge Δ (the sum of every factor) sets each side's expected goals from a base of 1.45 (home) / 1.25 (away).

Scoreline (Poisson)

P(k) = e^(−λ) · λ^k / k!

Independent Poisson draws on each xG (λ) produce every scoreline, then win/draw/win, BTTS, over/under, and projected scorers — all from one grid.

Player Index

100 · (0.30·Q + 0.30·O + 0.25·R + 0.15·M)

Squad Quality, projected Output, team deep Run, and Role/minutes — the transparent MVP race in the Player Rankings.

Worked example — Panama vs England

1Factors → edge

-669

Every factor sums to a net rating edge for England.

2Edge → expected goals

0.48 – 3.78

1.45 · e^(-669/600) · form = ×0.33

Base 1.45/1.25 goals, scaled by the edge and recent form.

3Poisson → scoreline

0–3

A Poisson grid on those goal rates yields every scoreline (12.7% most likely).

4Outcome

93%

2% Panama · 5% draw · 93% England.

The variables

Every factor Theo weighs

Team strength (Elo)

Theo strength model

weight ×1

Recent form

API-Football team form

weight ×0.6

Venue & host edge

Venue & altitude model

weight ×0.5

Squad quality

Theo squad ratings

weight ×0.4

Rest & congestion

Schedule (rest between matches)

weight ×0.4

Weather & heat

Theo weather (Open-Meteo)

weight ×0.3

FIFA ranking

FIFA/Coca-Cola World Ranking (fallback)

weight ×0.3

The data map

Where the signals come from

Strength & squadLive

Theo model

Elo-style ratings, squad quality, and FIFA rank — the always-on baseline.

Injuries & formLive

API-Football

Sidelined players, season stats, and form derived from real results.

Market consensusLive

The Odds API + Polymarket

De-vigged bookmaker and prediction-market probabilities.

Weather & heatLive

Open-Meteo

Matchday temperature, heat index, and WBGT heat-stress.

Web researchLive

Exa

Up-to-date, cited context surfaced through Ask Theo.

Monte Carlo

Title odds from 1,000 simulations

Theo plays the whole tournament out thousands of times, sampling each group scoreline and knockout winner from the model's own probabilities. The share of simulations a team lifts the trophy is its title probability.

1,000 simulationsseed 0x5eed · deterministic3/5 live signals

Results locked in through 2026-06-15 · live signals just now· bars are Theo's model odds, mkt = de-vigged market consensus.

Germany
20.3%mkt 2.5%
Argentina
18.2%mkt 2.4%
Netherlands
12.7%mkt 1.5%
France
11.2%mkt 2.2%
England
9.7%mkt 2.3%
Portugal
8.0%mkt 2.1%
USA
5.0%mkt 2.3%
Mexico
3.3%mkt 2.1%
Belgium
3.1%mkt 1.8%
Brazil
2.4%mkt 2.0%

Freshness

How often the title odds update

Recomputed on every page load, but the Monte Carlo is seeded (0x5eed) — identical inputs reproduce identical odds, so the numbers only move when the inputs move.
Inputs change when a match finishes (results locked in through 2026-06-15) or when live signals shift.
Live signals (injuries, form, scorers, market) refresh about every 5 min; betting odds about every 30 min. Last live-signal fetch: just now.
It's pull-based at request time — there is no background scheduler, so a value can be at most one refresh window stale.

Honesty

Calibration & limits

Predictions are probabilities, not certainties — every match card shows the full distribution, not just a single call.
When a feed isn't live (no key, or a market not yet posted), that factor falls back to a modeled prior and is labeled honestly.
Injury risk is a model prediction from workload, age, congestion, and heat — not a medical assessment.
The Monte Carlo is seeded, so the same inputs reproduce the same odds — results sharpen as real matches are played.