How to Read a Calibration Curve
A projection product can publish anything it likes about its own skill. A calibration curve is harder to dress up: it plots what the model claimed against what actually happened, one probability bucket at a time.
A projection product can publish anything it likes about its own skill. A calibration curve is harder to dress up: it plots what the model claimed against what actually happened, one probability bucket at a time.
A probability is incomplete until you know what happens when the player never takes the field. Void rules quietly redefine the event being estimated, and a model that ignores them is answering a question nobody asked.
Most arguments about player props are really arguments about vocabulary. One person says probability and means what a platform charges; another says it and means what a model estimated. This glossary defines each term narrowly and says what it does not mean.
Nineteen articles of theory compress into one working document. This is the checklist we would hand a friend who asked how to research props without fooling themselves: every step, the reason it exists, and the honest signals that a prop should be skipped entirely.
The same player, the same stat, the same number, and two completely different products. Esports props hide their most important detail in the scope of the line, and the most common research error in LoL and CS2 is analyzing one market while buying another.
Every soccer stat you can buy a line on flows through one narrow gate: time on the pitch. Before shots, passes, or tackles mean anything, research has to answer whether the player starts, and then whether they are still out there after the hour mark.
A fighter who lands six significant strikes a minute projects to ninety over a full fifteen minutes and to twenty if the fight ends midway through round one. Same fighter, same pace, wildly different props. Until you have an opinion about how long the fight lasts, you do not have an opinion about any volume stat in it.
Tennis is the rare sport where a prop can be worked out almost from first principles. Aces, games won, and total games all flow from two inputs, serve strength and return strength, pushed through a scoring system whose rules never change. The research problem is getting those two inputs right, and knowing which structural facts must be confirmed before any of it means anything.
In a league with shorter rotations and fewer nationally televised games, the variable that decides most WNBA props is not shooting form or matchup history. It is whether the player is on the floor, for how long, and in what role. Everything else is an adjustment to that answer.
The most expensive habit in hitter prop research is studying the batter and ignoring the batting order. Before talent, before matchup, before park, a hitter prop is a question about how many times a player walks to the plate.
Strikeout rate is one of the most stable skills in sports, which makes strikeout props a rare market where process genuinely compounds. The catch is that the stable rate sits behind an unstable quantity: how long the pitcher stays in the game.
An offer that hits 95 percent of the time sounds like the safest purchase on the board. Attach a reduced payout to it and it can quietly be one of the worst. Modified payout lines are where probability instinct fails, and where a small amount of arithmetic protects you.