Reading a Box Score for Prop Research

The box score was invented to answer a question that has nothing to do with research: what happened. It is a summary for someone who missed the game. Every design decision in it, from what gets a column to what gets printed in bold, serves that purpose, and the purpose is retrospective.

Prop research asks a different question. Not what happened, but what is likely to happen next, which requires separating the two ingredients a box score deliberately fuses together: how many chances a player got, and what he did with each one. Learning to pull those apart on sight is one of the highest leverage habits in this work, and it takes about ten minutes to learn and a season to stop forgetting.

Outcomes are printed large, opportunity is printed small

A player scored 24. That is an outcome. It is the number in the headline, the number in the recap, and the number a friend will quote at you. On its own it is nearly uninformative about next week, because at least three completely different games produce it: a heavy workload at ordinary efficiency, a normal workload at unusual efficiency, or a short stint where a few low probability shots fell.

The columns that resolve that ambiguity are the denominators. Minutes played. Plate appearances. Snaps and routes run. Touches, shot attempts, service games, rounds contested, possessions on the floor. These are the units of opportunity, and they carry two properties that make them far more valuable than the totals: they are mostly decided by role rather than by luck, and they are much more stable game to game than anything computed on top of them.

• Basketball: minutes first, then usage or shot attempts, then anything per attempt.
• Baseball: plate appearances and lineup slot for hitters, batters faced and pitch count for pitchers.
• Football: snaps, routes run, targets, and carries, all before any yardage number.
• Soccer: minutes and whether the player started, then shots and touches in the final third.
• Tennis: service games and total points contested, which scale with how long the match went.
• Esports: maps played and time on the server, since almost every stat is a rate over duration.

The pattern generalizes. Whenever you find yourself reading a total, ask what it was divided by and go find that number instead. If the source does not print it, the source is not built for research, which is a fact about the source rather than about the player.

How to read a big performance correctly

A standout line always has an explanation, and it is worth forcing yourself to name which one before drawing any conclusion. There are three candidates and they carry wildly different implications for the next game.

1. More opportunity. The player was on the field or the floor longer, or received more of the work, usually because someone else was absent or the game state extended it. Repeatable only if the cause repeats.
2. A better rate. Same opportunity, more production per chance. Sometimes real skill, more often the shortest sample in the box score doing what short samples do.
3. Luck on a small number of chances. Three deep shots fell instead of one. A ball found a gap. Nothing about the process moved at all.

The diagnostic is arithmetic and takes seconds. Compare the denominator to the player's typical denominator. If minutes or plate appearances or snaps were unusually high, most of the performance is explained by opportunity and you should be asking why, because that reason is what determines whether it happens again. If the denominator was ordinary and the total was not, the rate did the work, and now the question becomes how many chances the rate was computed over. A big rate over eight chances is a rounding artifact. A big rate over sixty chances is worth a second look.

One row is noise, a pattern in the denominators is signal

A single game log row is close to worthless in isolation, and this is not a claim about that particular game being weird. It is a claim about how much information one observation of a variable with real spread can carry, which is very little. Sample size in player prop analysis works through what it takes for a rate to say anything at all; the number of games required is usually larger than the number of games available.

But scanning a column of denominators is a genuinely different exercise from scanning a column of totals, and it becomes informative much faster. A player whose minutes have gone from the low twenties to the low thirties across five games has had his role changed, and five games is plenty to see that, because minutes are assigned rather than sampled. A player whose points have gone up over the same stretch with flat minutes has not yet demonstrated anything, because the totals are noisy in a way the minutes are not.

So read game logs column wise, not row wise. Put your finger on the opportunity column and look for a level shift. Then, separately, look at the rate column and ask whether it is drifting or just bouncing. Streaks and slumps covers why the bouncing looks so much like a trend to human eyes and why the eye is the wrong instrument for that judgment.

Four things a box score cannot tell you

Knowing what a source omits is as important as knowing what it contains, and box scores omit a specific and predictable set of things.

• Role changes below the resolution of the totals. A player can hold his snap count while losing a specific assignment inside it. The article on red zone usage walks through the extreme version: a goal line role measured in two or three snaps a game can disappear entirely without moving anything the box score prints.
• Health. A player who logged full minutes at reduced capacity looks identical on the page to a healthy one. Only reporting, and sometimes only the next few games, reveals it.
• Game context. Minutes accumulated in a blowout, a shootout, or an overtime period are not the same as minutes in a competitive regulation game, and the score state that produced them is not in the player rows.
• Whether the usage was sustainable. Opportunity created by an absence, an ejection, or a temporary lineup shuffle prints the same as opportunity created by a promotion. The box score records the amount and never the cause.

That last one is the most expensive omission in practice, because it is the one that most reliably produces a confident wrong projection. A player who saw a career high workload because two rotation pieces were out will show a beautiful recent line and a role that no longer exists by the time you act on it.

The filtered game log is a lie you tell yourself

Most tools let you filter a game log. Last ten games. Home only. Against this opponent. Since the coaching change. Every one of those filters is defensible in isolation, and the danger is not any single filter. It is that you get to keep trying them until one produces a subset that agrees with what you already suspected.

If you check five splits and report the one that supports your read, you have not found evidence. You have run five tests and published the winner, which is the same structural failure that makes a curated track record meaningless. Why we publish our misses is about that pattern at the level of a product; the filtered game log is the same pattern at the level of an afternoon.

The defense is procedural rather than statistical. Decide which split you are going to look at, and why, before you look at any of them. If the reason is causal, a new role, a different lineup, a surface change in tennis, then the split is a hypothesis and the data can test it. If the reason is that the unfiltered numbers were not cooperating, the split is a search for agreement and the data cannot test anything, because you have already decided what it will say.

Small sample splits deserve particular suspicion. Head to head history against one opponent, or performance in one venue, will almost always contain a subset small enough to swing dramatically. A big number over six games is what six games looks like. It is not a discovery.

A four step extraction routine

Do it in this order every time, and do not let your eye jump ahead to the totals, because once you have seen them you will read everything else as an explanation for them.

1. Denominators first. Minutes, plate appearances, snaps, touches, rounds. Look for the level, the recent trend, and any single game that is obviously out of pattern.
2. Rates second. Production per chance, computed over enough chances to mean something. Note how many chances each rate is built on, since that number governs how much weight it gets.
3. Context third. Score state, opponent, rest, absences, and anything that explains an unusual denominator. This is where you decide whether an opportunity spike was structural or temporary.
4. Totals last, and only as a check. If the totals surprise you after the first three steps, one of the first three steps was wrong or something happened that the box score does not record.

The last step is the one people find strangest, and it is the whole point. Totals are the output of everything above them. If you have read the denominators, the rates, and the context correctly, the total contains no additional information, and if it does contain a surprise, that surprise is a prompt to go looking rather than a fact to record. Hit rate versus projection makes the parallel argument about a player's history against a posted number: a count of past results is a summary of outcomes, and a projection has to be built from the inputs that produced them.

This is also, mechanically, what our engines do. They estimate opportunity, apply per chance rates, adjust for context, and produce a distribution rather than a single total. The difference is scale rather than method, and the reasoning behind each projection is published alongside it. Today's projections and the inputs behind them are on the signal board, and the graded record of how those projections resolved is in the Model Room.