Combinada Lab

Almost every betting tool multiplies the probabilities of each selection together. That is wrong whenever the selections come from the same match: P(over 2.5 ∧ both teams score) is not P(A)·P(B). Combinada Lab evaluates parlay legs over the model's joint distribution — a 4D score grid split by halves — so correlated legs are priced correctly instead of optimistically.

Underneath is a Dixon-Coles goal model with time decay, friendlies down-weighted, home advantage applied only at non-neutral venues, and a low-score correction; corners and cards come from a 100k-run Monte Carlo. Bookmaker margin is removed with Shin's method rather than naive normalisation, because Shin corrects the favourite-longshot bias that makes cheap value look real.

The part I care most about is the backtesting. It's strict walk-forward — at every date the model only knows earlier matches, and there are tests asserting that running a test cannot poison the future or change the past. It reports ROI at flat stake and fractional Kelly against real closing odds, plus a Brier comparison of model versus market. Often the honest answer is that the market is better, and the app says so.