What goes into the models
Four families of input feed every prediction. Results and ratings: every sport keeps a continuously updated strength rating per team or player, built from results, margins and opponent quality, the backbone that decades of sports modeling rest on. Matchup context: the variables that decide single games, like starting pitchers in MLB, confirmed goalies in the NHL, surface in tennis and stylistic matchups in the UFC. Market prices: odds from dozens of sportsbooks, treated as information in their own right, because the market is often the sharpest signal available. Situational data: rest, travel, schedule spots and lineup news, weighted by what actually moved results historically.
One model per sport, one shape of output
Sports price differently, so each gets its own model: goal-expectation simulations for soccer and hockey, rating-plus-matchup models for baseball, football and basketball, surface-adjusted player ratings for tennis, interaction models for MMA. What they share is the output contract: a probability for every outcome, an edge where a market price exists, and a confidence rating that says how stable that number is given the current inputs. You will never see a bare scoreline prediction here, because a scoreline without a probability is a guess wearing a costume.
Calibration: the standard we hold ourselves to
A prediction model should be judged the way weather forecasts are judged. When the model says 60%, that outcome should happen close to 60% of the time over a large sample. That property is called calibration, and it is the primary training target for every Oddify model. It is also why we publish probabilities instead of certainties: a well-calibrated 60% favorite loses four times in ten, and a model that hides that is lying to you. Resolved picks are logged and measured against results continuously, and models that drift get retrained.
How an edge is computed
Every betting line implies a probability: decimal odds of 2.00 imply 50%, minus the bookmaker's margin baked into the price. An edge is the gap between the model's probability and that implied probability. If the model makes a side 58% and the price implies 52%, the board shows a +6% edge. Two guardrails keep this honest. Edges are computed after stripping the bookmaker's margin, and implausibly large edges (beyond 20%) are treated as stale or broken price data rather than value, because in liquid markets a 40% edge is almost always an error, not a gift.
What the models cannot do
They cannot predict single games, only price them. They are late on news that breaks between updates: a scratched starter or a last-minute lineup change reaches the model when the data does, not when the tweet does, which is exactly what the confidence rating warns about. They are only as good as their inputs, and early-season or small-sample situations carry wide error bars that no amount of modeling removes. And they cannot beat the market everywhere: on most games the honest output is no edge, and the boards say so. Anyone promising daily locks in every sport is selling something else.
Frequently asked questions
Does Oddify guarantee winning picks?
No, and no honest model can. The models are built for calibration (probabilities that match reality over large samples) and for finding prices where the market disagrees with the numbers. Individual games remain uncertain by nature.
How often are the models updated?
Ratings update after every completed game, probabilities refresh on every model run through the day, and news-sensitive inputs like starting pitchers or confirmed goalies trigger re-pricing as they land.
Why do model and market sometimes disagree strongly?
Usually one of three reasons: the model knows something the market underweights, the market knows something the model has not ingested yet, or the price is stale. That is why very large edges are capped as suspect and why the confidence rating exists.
Are the models tracked against results?
Yes. Resolved predictions are logged and measured continuously, and that tracked record, not a backtest, is what drives which markets each sport's board actually shows.
See the models work on today's board
Every league page shows the model's live probabilities against the market, free to start.
Start free