WinEdge

Homepage
Head-to-Head Prediction Engine
Market vs Model Analysis (Alpha Detection Engine)
Alpha Signals & Decision Engine
Interactive Bracket Builder (User Picks & Model Comparison)
Live Squad Intelligence & News Feed
Travel Fatigue Analysis (WC2026 Model)
WinEdge- Social media flyer

What is WinEdge?

WinEdge is an AI-powered World Cup 2026 prediction engine that finds where betting markets are systematically wrong before the tournament starts. It combines a backtested Random Forest model, real-time Polymarket odds, travel fatigue modeling, and squad intelligence into a single deployed platform.

Try WinEdge Live | Live API

What Inspired Us

Every four years the World Cup generates billions in prediction markets. And every single tournament the market gets it wrong.

In 2022 Morocco were priced as a 200-1 longshot. They reached the semifinal, the first African nation in history to do so. The market missed it completely.

We asked a simple question:

What if you could find the next Morocco before the tournament starts?

That question became WinEdge.

How We Built It

We built the entire project in Zerve, from raw data to deployed API, in a single day.

Layer 1 — Data Pipeline

We loaded 49,287 historical international matches and filtered down to 2,257 competitive fixtures including World Cup, Euros, Copa America, African Cup of Nations, and Asian Cup from 2018 onwards. Friendlies and minor tournaments were excluded entirely.

Layer 2 — Feature Engineering

We engineered 21 features capturing team strength:

Neutral venue win rate and goal differential
Confederation-relative performance — a 60% win rate means something different in CONMEBOL vs OFC
Recency weighting: $ w_{2024-2026} = 3.0, \quad w_{2022-2023} = 2.0, \quad w_{2018-2021} = 1.0 $
Minimum match thresholds to prevent small-sample inflation

Layer 3 — The Model

We trained a Random Forest classifier with 300 trees and balanced class weighting on an 80/20 stratified split:

$$\text{Outcome} \in {0 = \text{loss},\ 1 = \text{draw},\ 2 = \text{win}}$$

Test accuracy: 55.97% — meaningful for a three-class football prediction problem where chance is 33%.

Layer 4 — 2022 Backtest

We retrained the model on pre-November 2022 data only and tested it against actual 2022 results.

Naive model directional accuracy: 31.2%
Improved model directional accuracy: 40.6%
Improvement: +9.4 percentage points

Most importantly, the model correctly flagged Morocco as undervalued at +3.0% edge before the tournament. Morocco finished 4th.

Layer 5 — Fatigue Model

We computed travel burden for all 48 teams using haversine distance across all 16 WC2026 host cities:

$$d = 2r \arcsin\left(\sqrt{\sin^2\left(\frac{\Delta\phi}{2}\right) + \cos\phi_1\cos\phi_2\sin^2\left(\frac{\Delta\lambda}{2}\right)}\right)$$

New Zealand travels 40,116 km — highest fatigue score in the tournament
Mexico travels 0 km — home advantage
CONCACAF teams receive a +3% home advantage bonus per match

Layer 6 — Live Market Comparison

We connected to the Polymarket API to pull live World Cup winner odds and compared them against our model probabilities.

Layer 7 — Alpha Signal Engine

We built a signal engine combining model probability, market odds, fatigue score, and injury data:

Signal	Condition
STRONG BUY	Adjusted edge > 2.5%, ratio > 3x
LEAN BUY	Adjusted edge > 1.0%, ratio > 1.5x
NEUTRAL	No significant divergence
FADE	Model < market by meaningful margin
STRONG FADE	Model < market, ratio < 0.5x

Layer 8 — Ask WinEdge: Natural Language AI Analyst

We integrated the Anthropic Claude API to build a natural language interface on top of the model. Users can ask anything about WC2026 in plain English and get data-backed answers grounded in our actual model outputs.

Example questions the system handles:

"What if Mbappe gets injured before the tournament?"
"Why is Brazil overvalued despite being a historic favorite?"
"Who is the biggest upset pick for 2026?"
"How was this model built?"

Claude receives the full model context — neutral venue win rates, market mispricings, fatigue scores, squad intelligence, and the 2022 backtest results — and returns sharp, specific, data-driven analysis in under 3 sentences.

This makes WinEdge the only World Cup prediction engine with a natural language interface grounded in a backtested statistical model. Every response ends with: "Powered by WinEdge — built on Zerve AI."

Layer 9 — Deployment

We deployed the full prediction engine as a live FastAPI on Zerve: GET /predict?team_a=Iran&team_b=England GET /signal?team=Iran GET /intel?team=Iran GET /fatigue GET /teams GET /signals/all We also built a production frontend WinEdge, calling the live Zerve API for real-time predictions. Features include:

Head-to-head probability engine
Market vs model mispricing table
Alpha signal cards
Team intel panels
Bracket builder with edge scoring
Squad intelligence and injury feed
Travel fatigue rankings

What We Found

The market systematically overprices brand-name teams and underprices consistent neutral-venue performers:

Iran: 57.7% neutral win rate, strongest in Asia. Market: 0.25%. Model: 4.0%. 21x undervalued.
Ivory Coast: 60% neutral win rate. Market: 0.45%. Model: 4.4%. 9.8x undervalued.
Brazil: only 41% neutral win rate since 2018. Market: 8.55%. Model: 2.0%. 4.3x overvalued.
Germany: 30% neutral win rate, worst among major European sides. Market: 9%. Model: 2.1%. 4.3x overvalued.

Challenges We Faced

The Small-Nation Bias Problem

Our first model gave artificially inflated probabilities to teams like Canada and Qatar because they had strong neutral venue records from a tiny number of matches. We fixed this with minimum match thresholds (8+ neutral matches required) and confederation normalization. This improved backtest accuracy by 9.4 percentage points.

The Zerve Deployment Environment

The FastAPI deployment runs in a read-only filesystem so we could not install packages at runtime. We solved this by using the FastAPI (Org) executor image which has dependencies pre-installed.

The Polymarket API

Odds are stored as JSON strings not native arrays, requiring careful parsing. We built a fallback to curated odds when the API returns incomplete data.

What We Learned

Football is genuinely unpredictable. 55.97% accuracy is the ceiling, not the floor. But relative edges are real and measurable. The market systematically misprices teams based on historical reputation rather than recent competitive form. Neutral venue data cuts through the noise.

The 2022 Morocco call was not luck. It was signal. And WinEdge is built to find that signal for 2026.

What's Next

The tournament starts June 11, 2026. WinEdge will update in real time as results come in. Every upset recalibrates every remaining team's probability before the market catches up.

We'll be ready.

Built entirely on Zerve.

Try WinEdge Live | Live API