A systematic betting operation on FIFA World Cup 2026 — a goal-scoring model prices every match, compares against real market prices, and bets only where it finds a genuine pricing edge on Kalshi and Polymarket.
Every model-recommended bet, logged before kickoff and published win or lose. Cumulative P&L is shown in dollars — the sample is small and the total is expected to be noisy; the discipline is the product.
Edge, not wins.
QuantBet doesn't try to pick winners — it hunts mispriced bets.A bet can be smart even if it loses, as long as the payout overcompensates for the risk. Think value investing: you're not buying the best team, you're buying whatever the market has priced wrong.
That's why the system is judged on expected value at bet time and closing-line value afterward — did we get a better price than where the market settled? Over a sample this small, win rate is mostly noise; CLV is the real report card.
One rule is absolute: the log above contains only bets the model itself recommended. Discretionary bets are tracked separately and excluded, so the record stays a clean answer to one question — is the model actually good?
Model edge vs. market price, per bet. Sizing is tiered fractional Kelly with a hard 3% bankroll cap — conservative by design, with a deliberately small platform balance as a guardrail against impulsive oversizing.
Six layers between a fixture and a stake.
Dixon-Coles bivariate Poisson
The statistical backbone — a goal-scoring model built for soccer, estimating each side's attack and defense strength with time-decayed match history.
LightGBM ensemble
A gradient-boosting layer on top of the Poisson engine, capturing interactions the parametric model can't express.
Isotonic calibration
Raw model outputs are recalibrated so a 60% prediction actually wins about 60% of the time — honesty enforced mathematically.
Market blending
Sharp market prices from The Odds API are folded in as an input — the model argues with the market, not in a vacuum.
Monte Carlo tournament sim
The full bracket simulated forward thousands of times for outright-winner and advancement markets.
Kelly-sized execution
Tiered fractional Kelly with a hard 3% bankroll cap per bet, executed on Kalshi (primary) and Polymarket. The logbet CLI handles sizing; log-result records outcomes.
Twenty-three bets is a start, not a sample.
Historical backtests
WC2018, WC2022, Euro2024, and Copa2024 replayed through the full pipeline — the real fix for the small-sample problem, prioritized once this tournament wraps.
Outright-winner report
A Monte Carlo tournament report for the World Cup outright market, sized with a separate, more conservative Kelly fraction for multi-round futures.
Combat sports expansion
UFC, boxing, and Muay Thai for more betting volume. The sizing, CLV, and logging infrastructure carries over; the prediction model gets rebuilt from scratch.