Methodology

The game model

For each matchup we compute a set of features describing both teams and feed them through a logistic-regression model whose coefficients are fit on thousands of past games. The features are league-specific — recent form, team strength and scoring differentials, home-field advantage, head-to-head history, rest, and the matchup’s key individuals (starting pitcher and bullpen in MLB, confirmed goalie in the NHL, quarterback in the NFL, and injury status in the NBA).

Critically, the model is point-in-time: when it evaluates a game it only ever uses information that existed beforethat game was played. This avoids “lookahead” leakage — the most common way sports models flatter themselves in backtests by accidentally peeking at the future — so the accuracy we report is the accuracy you can actually expect going forward.

Anchoring to the market

Sports betting markets are sharp. They aggregate enormous amounts of information — sharp money, late injury news, weather — that a games-only model can’t see. So after the model produces its own probability, we blend it with the de-vigged probability implied by current moneylines to produce the headline number you see on a prediction. The result is better calibrated than either source alone.

Calibrated confidence

A “65%” from our model is meant to be a real 65% — i.e., outcomes we tag at 65% should happen about 65% of the time over a large sample. We fit a calibration curve on held-out games so the numbers track reality rather than over- or under-stating confidence. Calibration is about being honest, not about being certain — see the limitations below.

The Edge Finder

An “edge” is the gap between the model’s own probability and the probability implied by the market price, after removing the sportsbook’s vig. We compute the edge from the model’s independent opinion — not from the market-blended number — because comparing a market-blended probability back to the market would be circular. When an edge clears a meaningful threshold, we also show a fractional Kelly stake suggestion, which sizes a bet in proportion to its edge while controlling risk. An edge is an opinion about value — never a guarantee of profit.

Our tracked record

We log every prediction before the game starts, then score it once the game finishes. Nothing is edited after the fact, and we report both win accuracy and the Brier score — a standard measure of how well-calibrated a set of probabilities is (lower is better). This is the number that matters: anyone can claim a hot streak; we publish the running record so you can judge for yourself.

(The record accumulates forward — today’s predictions are scored after tonight’s games finalize.)

Beyond win probabilities

• Run/score totals — a separate regression for over/under environments, including park and weather context.
• Power rankings — a transparent z-score blend of win rate, scoring differential, and recent form.
• Title odds & projected wins — a 10,000-iteration Monte Carlo simulation of the remaining schedule.
• DFS projections — per-player point projections built from recent usage, matchup, and the game’s expected scoring environment.

Honest limitations

• Sports are high-variance. A well-calibrated 60% pick still loses 40% of the time — that’s not a broken model, that’s the math.
• Betting markets are efficient. A consistent, large edge over the closing line is rare; we’d rather show you a small honest edge than a fabricated large one.
• Early in a season, samples are small and the models lean more on priors and the market.
• Data can be late or wrong (lineups, injuries). We mitigate this but can’t eliminate it.
• Nothing here is betting or financial advice, and none of it guarantees an outcome.

Data sources

We build on official league data (MLB StatsAPI, the NHL API), ESPN for scores and schedules, The Odds API for market lines, and Open-Meteo for weather. Team names and logos belong to their respective leagues and are used for identification only.