A Player Prop Glossary: Precise Definitions for Research

Most disagreements about player props are vocabulary problems wearing an analytical costume. One person says probability and means the number a platform's price implies. Another says it and means what a simulation produced. A third says edge and means the distance between a projection and a line, which is a different quantity again. The conversation goes nowhere because nobody was talking about the same object.

What follows is a working glossary, organized in the order research actually encounters these ideas: the offer you are looking at, the numbers attached to it, the tools used to evaluate it, and the rules that decide what it becomes afterward. Each entry says what the term means and, where confusion is common, what it specifically does not mean. If you want the process these terms serve rather than the terms themselves, start with the research guide.

Group one: the offer

Everything begins with the thing a platform is actually selling. Get this layer wrong and no modeling downstream can rescue the analysis, because you will be evaluating a product that is not for sale.

• Prop. Short for proposition. An offer settled on a single participant's statistical output rather than on which side wins. A player prop is not a wager on the game result, and its correlation with the game result varies enormously by sport and stat.
• Line. The threshold number a platform posts for a stat, such as 6.5 strikeouts. It is a claim about where the platform is willing to split the outcome distribution in two. It is not a prediction of the player's exact total, and it is not the same thing across two platforms that happen to print the same digits under different rules.
• Over and under. The two sides of a threshold. Over means the settled statistic finishes above the line; under means below. Not every offer sells both. Some fantasy style products list only the over, in which case disliking the over means passing, not taking an under that does not exist.
• Over only offer. An offer where the platform sells one direction. Disagreement produces a pass. Treating a disliked over only offer as an implicit under is the single most common way research produces a recommendation nobody could act on.
• Alternate line. A second threshold on the same stat for the same player, sold alongside the standard one. An alternate line is not a correction of the standard line and does not imply it was wrong. It is a different split of the same distribution.
• Modified payout line. An alternate line whose payout also changes: a discounted line moves the threshold in your favor and pays less, a boosted line moves it against you and pays more. Some apps give these nicknames. The nickname is branding; the mechanism is a price attached to a probability. Multipliers and eligibility rules differ by platform and change over time, so the platform's published rules are the authority.
• Slip or entry. A grouping of picks that must all settle correctly, or reach some threshold count, for the entry to return anything. A slip is not the sum of its picks; its behavior depends on how the picks relate to each other.

Group two: the numbers

Three quantities get conflated constantly, and the conflation is expensive because they can point different directions on the same offer.

• Projection. A model's estimate of a quantity, such as an expected total or a probability that a threshold is cleared. It is an estimate with error around it, not a forecast of what will happen. A projection is not a line and does not become one by being confident.
• Posted line. What a platform sells, defined above. The distinction from projection matters because research compares them; if you use the words interchangeably you have quietly assumed the answer.
• Market reference. A fair probability derived from markets after removing the operator's margin, usually by combining several books that trade the same event at meaningful volume. A market reference is not simply another platform's number, and it does not exist for every sport or every stat. Where none exists, honest analysis is model versus line comparison and should say so.
• Implied probability. The probability a posted price appears to represent before removing margin. Because margin is included, the implied probabilities of both sides of a market sum to more than one hundred percent. Implied probability is therefore not a fair probability and is systematically too high on both sides at once.
• Fair probability. What remains after the margin is stripped out, so that the sides sum to one hundred percent. Fair probability is an estimate of the market's opinion, not a measurement of truth. Sharp markets are frequently wrong; they are just wrong less predictably than most alternatives.
• Vig, also called the margin or the hold. The operator's built in cut, expressed as the gap between the sum of implied probabilities and one hundred percent. Vig is not a fee charged after the fact. It is priced into the numbers before you ever see them, which is why comparing raw implied probabilities across products is meaningless.
• Edge. The difference between a model's probability for a side and a reference probability for the same side. Call it a market edge only when a real market reference exists. The distance between a projection and a posted line without any reference behind it is a disagreement, which may be informative and may simply mean the model missed news.
• Expected value. Probability of a result multiplied by what that result returns, summed across outcomes. Expected value is the only lens that ranks offers correctly when payouts differ between them, which is exactly the case on modified payout lines. It is not a prediction of any single outcome and does not become one over small samples.

Group three: evaluation

These are the terms used to judge whether a projection process is any good. They are the terms most often used loosely, because loose usage flatters whoever is publishing the record.

• Hit rate. The share of decided recommendations that settled correctly. Hit rate is not accuracy and not skill. It depends entirely on which offers were selected and at what payouts, so a high hit rate on structurally easy discounted lines can describe a losing process. Compare hit rates only within a matched class of offers.
• Calibration. Whether stated probabilities match observed frequencies: of everything a model called 60 percent, roughly 60 percent should happen. Calibration) is not the same as being right often. A model can be perfectly calibrated and completely useless, if it says 50 percent about everything.
• Discrimination, sometimes called sharpness. The ability to separate likely from unlikely events, pushing probabilities away from the base rate when justified. A model needs calibration and discrimination together. Either alone is a partial claim.
• Brier score. A single number scoring probabilistic forecasts, lower being better, which blends calibration and discrimination. Useful for comparing versions of the same system on the same events. Not comparable across sports, markets, or eras with different base rates.
• Variance. The spread of outcomes around an expectation. Variance is not error and not bad luck. It is a structural property of the thing being predicted, and a correct 58 percent estimate is expected to lose 42 times in 100.
• Sample size. How many decided observations a claim rests on. Small samples produce dramatic hit rates for free. The right instinct is that a number's volatility should scale with how small its sample is, and that a claim from twenty results is closer to a rumor than a measurement.
• Shrinkage, also called regression toward a prior. Pulling a small sample estimate toward a broader baseline in proportion to how thin the evidence is. Shrinkage is not pessimism. It is the correct treatment of an estimate whose noise is large relative to its signal, and skipping it is how a hot ten game stretch becomes a projection.
• Correlation. The tendency of two outcomes to move together, positively or negatively. Correlation is not causation and, more practically, is not a reason to stack: correlated picks raise both the chance everything hits and the chance everything misses, which changes the risk shape of an entry rather than improving it.
• Closing line value. The comparison between the number you engaged at and where that number finished before the event. It is a diagnostic about timing and information, not a result, and it only means anything in bulk.

Two of these terms deserve extra suspicion whenever you meet them in marketing. Hit rate is the one most often quoted without saying which offers it counts. Edge is the one most often quoted without saying what it is measured against. Both questions are fair to ask of any tool, including ours in the Model Room, and the deeper treatment of the difference between hit rate and projection quality lives in that comparison article.

Group four: settlement

Settlement is where an offer becomes a result, and where research most often discovers it was estimating the wrong quantity all along.

• Settlement. The application of a platform's rules to official results to decide what a pick becomes. Settlement is not the box score. Which stat source counts, whether overtime or extra time is included, and what happens to partial participation are all rules choices that vary by platform.
• Push. An exact tie between the settled statistic and a whole number line, such as 2 hits against a line of 2. Platforms handle pushes differently and some avoid them by posting half numbers. A push is not a loss and not a win, which means probability models must carry the tie mass explicitly rather than assuming the under is one minus the over.
• Void. A pick removed from settlement, typically returning the stake or being dropped from an entry. Void is not a loss. Its existence is why published probabilities on many products are conditional rather than absolute.
• DNP. Did not play. The participant was available in principle but recorded no appearance. On most player prop products this triggers a void rather than a settled under, which means the probability you should care about is the probability given that the player appears.
• Scratch or late withdrawal. Removal from the participant list before the event starts. Usually simpler than a DNP because the offer is normally pulled, but timing matters, and an entry submitted before the news may be treated differently from one submitted after it.
• No play. Not a platform term at all, but a research output: the honest conclusion that an offer does not merit action. A no play is not a weak lean. In a probabilistic process it should be the most common conclusion, because most posted lines are placed roughly where they belong.

Using the vocabulary

Vocabulary discipline is not pedantry here; it is the cheapest available error correction. Say projection where you mean projection and market reference where you mean market reference, and several bad conclusions become impossible to state. You cannot claim a market edge in a sport where no reference exists. You cannot rank a discounted line above a standard one on probability alone, because expected value is now in the sentence demanding a payout.

It also changes how you read other people's claims. A tool that publishes edge without naming its reference, or a record that pools modified payout offers with standard ones, is not necessarily dishonest, but it is making a claim less precise than the one it appears to make. The modified payout article works through why that particular pooling flatters a record, and the model, line, and market piece works through what a reference has to be before edge is a legitimate word.

None of these definitions make outcomes predictable. They make claims checkable, which is a smaller and more useful thing. Keep the activity recreational, set limits before a slate rather than during it, and if it stops being recreational, help is available at 1 800 GAMBLER and through the National Council on Problem Gambling. To see the vocabulary attached to live offers, projections, and honest no play labels, open the Signal Board.