Public method
How the forecasts work
Predict Flow is an independent model next to the market — not a black-box “supercomputer.” The public pages show remaining-season tables and match scorelines. The paid slate adds edge, Kelly fractions, and the full daily card.
Soccer match model
Team attack and defense rates come from season xG and goals. Those rates feed a Dixon–Coles Poisson grid (ρ = −0.11) for 1X2, BTTS, and over/under 2.5. Table position and last-five form blend in, then a logit mix toward book odds when we have a real market.
Players change λ on the explained scoreline grid. A typical XI is the 11 highest-minute players with a 1-4-4-2 shape. Confirmed or predicted starters scale team λ by (XI weight / typical-XI weight), clamped 0.72–1.28. If no lineup is in yet, we use the minutes prior — never a silent 1.0 that hides absences.
The billed 1X2 pick stays on team Dixon–Coles until lineup λ wins a holdout. Scorelines, BTTS, and remaining-season tables already use the XI-adjusted grid.
Season simulations
Remaining league fixtures are sampled 15,000 times from the same Dixon–Coles grid. Played results stay locked. We publish title, top-four (or MLS playoff), and relegation percentages only when a competition has team stats and enough remaining matches. Thin leagues stay on the slate as low-confidence model rows.
MLB remaining-season odds blend current win percentage with team OPS and ERA from the same lineup engine as the daily picks, then simulate the rest of the schedule for division, wild-card, pennant, and World Series. NBA and NFL playoff tables appear when those seasons have remaining games; injuries are minutes- or usage-weighted rather than a headcount.
Frozen before kickoff
Accuracy is graded from the pre-match ledger. We do not rewrite a pick after the whistle. Public hit rates are winner-pick accuracy with sample size — not betting profit. Margin slices (≥5% / ≥8% from a coin flip) show how the model does on stronger favorites.