Glossary
128 terms. Each has a one-line definition, a plain-English explanation, an example, related terms and the lesson or guide that teaches it.
A range that, under stated assumptions, contains the true value with a given frequency (e.g. 95%).
The plausible range around a measured number. Small samples produce wide ranges.
The full set of possible outcomes and the probability of each.
Not just the most likely result — every result and how often it happens. Over/under probabilities come from the distribution, not the average.
The compact description of an event’s current situation that the next moments depend on.
Score, time and the sport-specific situation — outs and runners, down and distance, strength, round.
The expected count parameter of a Poisson distribution.
The average number of events you expect in the window. Everything in a Poisson model flows from it — and it is always an estimate.
How much a single event can change the outcome probability from the current state.
How much this moment matters. Late, close situations are high leverage.
A win probability updated on the current game state.
The chance of winning from here, given everything that has happened.
Estimating a probability by running a random model many times and counting outcomes.
Play the game (or fight, or season of bets) thousands of times on a computer and see how often each result happens.
The fraction of simulations that ended in a given outcome.
How often something happened in the simulation — only as reliable as the model that produced it.
When observed variance exceeds what a Poisson model with the same mean allows.
Real outcomes are more spread out than the simple model assumes, often because the rate itself changes from game to game.
A distribution for counts of independent events in a fixed window at a steady average rate λ: P(k) = λᵏe^(−λ)/k!.
A way to turn "expected 0.35 home runs" into the chance of 0, 1, 2… — valid when events are rare, independent and the rate is steady.
A probability estimated before the event starts, averaging over every way it could unfold.
The before-the-start number. It is graded on its own; live swings do not make it wrong.
A number from 0 to 1 describing how often an outcome would happen across many identical situations.
How likely something is, expressed as a share of repeated tries. It describes uncertainty; it does not predict a single result.
A single summary of a forecast distribution, usually its mean or median.
The middle of the range of likely outcomes. Useful for orientation; not directly comparable with a price.
The tendency of extreme results from small samples to be followed by results closer to the underlying average.
Hot and cold streaks are partly luck, and luck does not persist, so later results drift back.
The number of observations behind an estimate.
How much evidence a number rests on. Most sports records are far too short to prove much on their own.
The typical sampling error of an estimate; for a rate, √(p(1−p)/n).
How much a measured rate would wobble if you re-ran the same number of trials.
A measure of how spread out outcomes are around their average.
How much results bounce around. High variance means short-run results tell you little about the long-run average.
Odds anchored on $100: +X is profit on a $100 stake; −X is the stake needed to profit $100.
The plus/minus format. +150 wins $150 on $100; −150 risks $150 to win $100.
The money set aside for betting, separate from everything else.
A fixed budget. Sizing decisions are percentages of it.
The win rate at which a price returns exactly zero over the long run; equal to its implied probability.
How often you need to win at a given price just to not lose money.
The last price available before an event starts.
The market’s final answer, after lineups, injuries and news are in.
How much better or worse the price you took was than the closing price.
Did you beat the market’s final estimate? A faster skill signal than win-loss, and best measured against the no-vig close.
The degree to which two outcomes tend to happen together.
If one result makes another more likely, multiplying their probabilities gives the wrong answer.
Total return per unit staked, including the stake.
The easiest format for math: profit per unit is decimal − 1 and implied probability is 1 / decimal.
A decline from a bankroll’s peak to a later low.
How far you fall from your best point during a losing stretch.
The difference between an estimated probability and the (preferably no-vig) market probability.
How far your number is from the market’s. A hypothesis about accuracy, not a guaranteed profit.
The probability-weighted average result: p × profit − (1 − p) × stake.
What a bet returns on average if your probability is right. It is only as good as that probability.
Prices implied by probabilities that sum to 100%, usually estimated by removing the vig.
The price with the book’s margin taken out — a better estimate of what the market actually believes.
Staking the same amount on every bet regardless of perceived edge.
Same size every time. Simple, auditable and forgiving of overconfident probabilities.
The share of total money a book keeps on perfectly balanced action: 1 − 1 / (sum of implied probabilities).
The book’s theoretical profit margin. Same idea as overround, measured differently.
The probability at which a price breaks even: 1 / decimal odds. It still contains the book’s margin.
The probability the price requires you to believe. Not the book’s honest estimate — the vig is still inside it.
The stake fraction f* = (p·b − q)/b that maximises long-run growth when p is known exactly.
A sizing formula that bets more when the edge is bigger. It assumes your probability is right — which it never exactly is — so most users take a fraction of it.
The amount by which a market’s implied probabilities exceed 100%.
How far the prices overshoot a fair 100% book.
A combined bet that wins only if every leg wins; decimal prices multiply.
Several bets chained together. Probabilities multiply too, and so does the book’s margin.
A market on a specific outcome within an event rather than its final result.
A price on one slice of a game: a player’s strikeouts, yards or points, or whether a fight goes the distance.
Net profit divided by total amount staked.
Profit per dollar risked. Always check the staking standard behind a published ROI.
The margin a book builds into its prices so both sides’ implied probabilities sum to more than 100%.
The book’s cut, hidden inside the odds rather than charged as a fee.
Evaluating a model on historical data as if it had been used at the time.
A replay of the past. Only trustworthy with time-respecting splits and no leakage.
A simple reference forecast a model must beat, such as 50/50 or the historical win rate.
The “could a simpler rule do this?” test.
The mean squared difference between probability forecasts and outcomes (1 or 0). Lower is better.
Grades probabilities, not picks. Always forecasting 50% scores 0.25 — the no-skill line.
Agreement between forecast probabilities and observed frequencies across many forecasts.
When a calibrated model says 70%, it happens about 70% of the time.
The weighted average gap between forecast probability and observed frequency across buckets.
One number for how far off the probabilities are, on average.
Data withheld from fitting and tuning, used once for a final performance estimate.
The exam the model never studied for.
Information from the outcome, or from after the prediction time, entering a model’s inputs.
The model accidentally peeking at the answer. Produces great backtests and bad real results.
The negative average log-probability assigned to what actually happened. Lower is better.
Like Brier, but punishes confident misses far harder. A coin flip scores ln 2 ≈ 0.693.
Fitting noise in training data so performance on new data is worse than on the data used to build the model.
A model that memorised the past instead of learning from it.
A scoring rule whose expected score is best when the forecaster reports their true belief.
A grading system you cannot game by shading probabilities. Brier and log loss are proper; win rate is not.
A plot of observed frequency against forecast probability, by bucket.
The picture of calibration: dots on the diagonal are calibrated.
Repeatedly fitting on all data up to a point in time and scoring only the next period.
Testing the way the model is actually used: only the past predicts the future.
A Statcast batted-ball classification combining high exit velocity (98+ mph) with a launch angle in a band that widens as exit velocity rises.
The best kind of contact: hit hard at a good angle. Barrel rate is one of the strongest indicators of home run power.
The combination of occupied bases and number of outs (24 states).
The situation in the inning. It sets how many runs the rest of the inning is likely to produce.
Called strikes plus whiffs divided by total pitches.
How often a pitcher gets a strike without the ball being put in play. A strong strikeout indicator.
The speed of the ball off the bat, in mph.
How hard the ball was hit. Harder contact produces more hits and more power on average.
An index of how a ballpark affects a stat relative to league average (100 = neutral).
Some parks help hitters, some suppress them. It is a modest adjustment, not a switch.
A player’s performance split by the handedness of the opponent (batter vs RHP/LHP).
Most hitters do better against opposite-handed pitchers. Splits need big samples before they mean much.
The average runs scored from a base-out state to the end of the inning.
How many runs this situation is worth, on average.
Expected weighted on-base average, estimated from exit velocity and launch angle (and sprint speed on some batted balls), plus strikeouts and walks.
How well a hitter “should” have done given the quality of contact, stripping out defence and luck on balls in play.
The distance a pass travels past the line of scrimmage to the catch point or intended catch point.
How far downfield a player is targeted — a measure of role and ceiling.
A market on whether a player scores at least one touchdown in the game.
A yes/no scoring market. Low-count, high-variance, and often priced with heavy margin.
The current down and yards needed for a first down.
The core of football’s game state. 3rd-and-1 and 3rd-and-12 are very different situations.
How the score and clock shape play-calling as a game unfolds.
Teams that lead run more; teams that trail pass more. It shifts player volume.
A receiver’s targets divided by the team’s pass attempts (or team targets).
How much of the passing game goes to a player. Opportunity drives receiving production.
A player’s share of team opportunities (snaps, routes, targets, carries, possessions).
How involved a player is. Changes in usage usually matter more than changes in efficiency.
A single score built from weighted components.
Several stats blended into one number. Read the components and weights, not just the total.
Time on court; the main driver of counting stats.
No minutes, no stats. Minutes restrictions and rotation changes move props more than matchups do.
Possessions per 48 minutes (NBA) or per 40 minutes (WNBA).
How fast a game is played. More possessions mean more chances for every stat.
The share of a reference group that a value exceeds.
Where a player ranks relative to peers, from 0 to 100.
A team’s continuous control of the ball, ending with a score, turnover or defensive rebound.
The basic unit of basketball: each team gets roughly the same number per game.
The probability a shot becomes a goal given its characteristics, summed over shots.
How many goals a set of chances “should” have produced. Needs validation before use — heuristics are not xG. PropBetEdge Soccer has not built its own xG; where a source’s xG appears, it is labelled as source-supplied.
Saves divided by shots on goal faced.
How often a goalie stops shots. Noisy over small samples.
Shot attempts that would have entered the net without a save.
Shots the goalie had to stop (or didn’t). Blocked and missed shots do not count.
The number of skaters per side (5v5, 5v4 power play, 6v5 empty net).
Hockey’s key live condition. Scoring rates change dramatically on the power play.
A Poisson goals model that adjusts the probabilities of low scores (0–0, 1–0, 0–1, 1–1) and weights recent matches more heavily.
Two Poissons, one per team, with a correction where independent Poissons are known to be wrong. A standard, transparent starting point for soccer.
A completed pass into or within the attacking third of the pitch.
A measure of how often a team or player moves the ball into dangerous territory. It describes territory, not possession.
The average improvement in results for the home team, all else equal.
Playing at home helps, modestly. Its size varies by league and season, so models estimate it rather than assume it.
A pass that directly leads to a teammate’s shot.
The pass right before a shot — a volume measure of chance creation, whether or not the shot scored.
A count normalised to 90 minutes: count × 90 / minutes played.
Lets players with different minutes be compared. Tiny samples produce extreme rates, so a minutes gate is applied before ranking.
The current scoreline from one team’s perspective — leading, level or trailing — and by how much.
Teams play differently depending on the score: leaders sit deeper, trailing sides push forward. Match totals mix skill with these effects.
One team’s fraction of all completed passes in a match.
Who completed more of the passing. It is NOT possession percentage: without a possession clock, the event data cannot measure time on the ball.
A shot that would have entered the goal without a save, or that scored.
An attempt the goalkeeper had to deal with. Blocked shots and shots wide or over do not count.
A market with three outcomes: home win (1), draw (X), away win (2).
Soccer’s draw makes the result a three-outcome question. All three implied probabilities share the book’s margin.
Aces divided by service points.
How often a serve wins the point outright, untouched. A serve weapon; strongly affected by surface.
A point on which the returner wins the game if they win the point.
A chance to break serve. Break points saved and break-point conversion measure who wins these high-leverage points — on small samples.
The share of return games a player wins (breaks of serve).
How often a player takes the opponent’s serve — usually what decides a set.
The final possible set of a match: the third in best-of-three, the fifth in best-of-five.
Winner-take-all. Deciding-set records are small samples and are often over-read as “clutch”.
Double faults divided by service points.
How often the server gives the point away. Read it alongside ace rate: aggressive servers trade one for the other.
The share of service games a player wins.
How reliably a player keeps serve. Hold rate is a game-level rate; it is not the same thing as match win probability.
A match ended because a player stopped during play, usually through injury or illness.
A result without a completed match. Records and statistics treat retirements differently from completed results.
The share of points a player wins on the opponent’s serve, often split into first- and second-serve returns.
The core return number — the mirror of the opponent’s service points won, and the engine of breaks.
The share of points a player wins when serving, across first and second serves.
The core serve number. Small changes compound into large changes in hold rate.
A match won without losing a set (2–0 or 3–0).
A one-sided match by sets — though individual sets can still be close.
A player’s statistics separated by court surface: hard, clay or grass.
Surfaces change how much the serve is worth, so splits are informative — but each split is a smaller sample.
A game played at 6–6 in a set (rules vary by event) to decide the set, usually first to 7 points by two.
The shoot-out when neither player breaks. Short, high-variance, and dominated by serve.
A match awarded because a player withdrew before it started.
No ball was struck. It advances the opponent but says nothing about either player’s level.
The standard scoring system: the round winner gets 10 points, the loser 9 or fewer.
How judges score rounds. Close rounds are 10–9; dominant rounds can be 10–8.
A bout that reaches the judges’ scorecards without a finish.
The fight lasts all scheduled rounds. Includes draws.
The probability a bout ends by a given method (KO/TKO, submission, decision).
How it might end, not just who wins. Harder to predict than the winner.
A count normalised to 15 minutes of fight time: count × 900 / observed seconds.
Makes fighters with different amounts of cage time comparable.
The number of rounds a bout is booked for (UFC: 3 or 5; boxing: typically 4–12).
How long a fight can last. It changes distance and finish probabilities.
A decision in which two judges score for one fighter and one judge for the other.
The judges disagreed. Close fights produce splits, and judging varies by official.
Significant strikes landed minus significant strikes absorbed, usually per minute or per 15.
Whether a fighter out-lands opponents on average.
A fighter pulling out of a scheduled bout.
A major context change: the bout may be cancelled or a replacement booked.
The difference between the highest price a buyer bids and the lowest price a seller asks.
The gap you cross to trade immediately. Wide spreads mean the headline number is less precise than it looks.
A contract that pays a fixed amount if a defined event happens and nothing if it does not; binary contracts commonly settle at $1 (YES) or $0 (NO).
A written yes/no question with money attached. Its price in cents reads naturally as a market-implied probability.
The tendency, once an outcome is known, to believe it was more predictable than it was.
After the fact, everything looks obvious. Judge a forecast by what was knowable when it was made.
The quantity available at each price level of an order book.
How much can change hands before the price moves. Thin depth means the best price only holds for small size.
A set of contracts on one event where at most one outcome can resolve YES, such as a ladder of rate-decision buckets.
Several buckets that cannot all be right. Read their prices together, not one at a time.
The named authority and published value that decide how an event contract resolves, as written in its rules.
Who decides, and from which number. Not the headline, not a forecast app — the source the rules name.
The final payout of a contract once the venue determines the outcome under its rules, which can include special outcomes such as a tie or a fair-price resolution.
When the contract is closed out and paid. Ties, delays and corrections can make settlement more than a simple yes or no.
The requirement that a forecast and the market price it is compared with were both knowable at the same moment, before the contract locked.
Compare numbers from the same moment. A later price may already contain news the forecast never saw.
Whether, and how much, a player or fighter is expected to participate.
Who is actually playing — usually the biggest single input to any projection.
The network’s rule for which bouts are on a card: only confirmed evidence removes a bout.
A reported withdrawal is a warning; a confirmed one removes the bout. No guessed reasons.
A label (e.g. Lean / Medium / High) describing the probability band and the evidence behind a call.
How much evidence supports the number — not a promise about the result.
A label describing whether enough verified data exists for a reliable output (e.g. FULL / LIMITED / INSUFFICIENT DATA).
Whether there is enough history to trust the calculation.
The ordered record of sourced match events — shots, passes, fouls, cards, substitutions — each with a minute and, for some, a location.
What happened, in order. Soccer PBEcast plays it back; it is not player tracking.
PropBetEdge’s description of a fighter’s official record as normalised rates with sample-size tiers.
What the record shows — not a prediction and not a grade.
The moment a call is frozen before an event; it cannot be changed afterwards.
The time-stamp that proves the pick came before the result.
A published document describing a model: what it predicts, training data window, out-of-sample metrics, baselines and limitations.
The model’s spec sheet and report card. Read it before trusting the numbers.
A designed refusal to publish a probability or pick when the inputs would make it unreliable.
The model saying “not this one” — debuts, missing features, too close to call.
The model probability minus the de-vigged market probability, in percentage points.
How far a PropBetEdge model disagrees with the market after removing the book margin. Disagreement is not proven value.
PropBetEdge Tennis’s Elo-style rating, updated from match results.
A results-based strength number. Matchup win probabilities from it are published only where a tour’s rating has beaten a ranking baseline in backtests.
PropBetEdge Soccer’s descriptive player profile: per-90 rates ranked against the same competition-season, behind a 450-minute gate.
Where a player ranks among qualified peers in the same league and season. Descriptive, not a forecast.
Pre-set criteria a model must pass before its outputs are published as official.
The bar a model clears before its picks count. Until then it runs in tracking or shadow.
A model that runs and is evaluated privately without publishing its outputs.
A model on probation: measured, not shown.
PropBetEdge Tennis’s descriptive player profile: Match DNA from results and technical DNA from serve and return statistics, with same-tour percentiles and sample gates.
What a player’s record shows, ranked against qualified players on the same tour. Descriptive, not a win probability.
The graded history of locked calls.
Every call, graded, never rewritten. Read it with sample size and prices in mind.
A graded record of real pregame decisions kept separate from the official record while a model earns promotion.
A verified trial record — never added to the official one.
A published WNBA composite of production percentile, win rate, winning-output share and court share.
A descriptive ranking of winning production — not a causal estimate and not a pick.
No terms match. Try the network search.
