Why a Model Probability Is Not a Recommendation

A count model does more than produce an average. Given its inputs, it describes a whole range of possible outcomes: how likely zero is, how likely one is, how likely two or more is, and so on. From that range it can state the chance the outcome finishes above a particular number.

As an illustration, a model might say that the chance a player finishes above 1.5 of something is 47%. That is a statement about the model’s own view of the range of outcomes, frozen at the time the projection was made. It is not a forecast that the outcome will land on one side, and it says nothing on its own about any price.

FDI models estimate these probabilities at a small, fixed set of thresholds chosen before the season — for example 0.5, 1.5 and 2.5 for a low count, or 1.5 through 5.5 for a larger one. Those thresholds are part of the recorded prediction. They are not recomputed later for whatever number happens to be convenient.

A sportsbook line is one number. If the current line equals one of the model’s thresholds exactly — the line is 1.5 and the model estimated the chance of finishing above 1.5 — the two can sit side by side honestly. The probability answers the question the line poses.

If the line is anything else, FDI shows no probability at all. It does not interpolate between two thresholds, it does not extend the range past the last one, and it does not recalculate the model on the fly. Each of those would produce a number that looks precise but that the recorded model never actually stated.

• Current line equals a modeled threshold for the same player → the frozen probability for that threshold is shown.
• Current line falls between thresholds or outside them → no probability; the page says it is shown only at a modeled threshold.
• The line is not current, or belongs to a different player → no probability.
• No projection has been published yet → no probability treatment of any kind.

See how the dossier presents it in the sample

Even a probability that sits exactly on the line leaves the most important questions open. The line is offered at a price on each side, and those prices include the book’s margin. At −110 on both sides, a side needs to be right about 52.4% of the time across many repetitions just to break even, because 110 ÷ 210 ≈ 0.524. A model saying 51% does not clear that; a model saying 55% might or might not, depending on how well calibrated it has proved to be.

Calibration is the second open question. A stated 55% is only meaningful if, across many predictions, outcomes the model called 55% happen close to 55% of the time. That can only be learned by recording predictions before games and checking them afterward, over a meaningful sample. A single probability cannot establish it.

For those reasons FDI labels a model probability as model context and nothing more. It is not a recommendation, it is not ranked, and it is never converted into a suggested side. What a member does with context is their own decision.

Model vs market: projections, lines and prices

A model probability is only as complete as the model’s inputs. It is frozen at a fixed time before kickoff, so news that arrives afterward — a late scratch, a change in role, weather that turns — is not in it. The market may already reflect some of that information when the model cannot.

It also assumes the player takes part in the game in a normal role. If that assumption fails, the probability was describing a different game from the one that was played. FDI records those cases separately rather than folding them into a miss, but a pregame reader cannot know in advance which games they will be.

• Information that arrives after the projection was frozen.
• How much of the same information the market has already priced.
• Whether the player’s role changes before or during the game.
• Whether the model’s probabilities have proved well calibrated in live use.

Because an average alone hides the shape of the outcome. Two projections with the same average can have very different chances of finishing above a line: a count that is usually zero or one behaves differently from one that is spread evenly across a wide range. The threshold probability makes that shape visible at the one number where it matters.

Shown carefully — only at an exact threshold, only against a current line for the same player, and labelled as model context — it helps a reader understand the model’s view. Presented any other way it would invite more confidence than the evidence supports, which is the opposite of what FDI is for. When there is not enough information to say anything useful, FDI says that instead.

Why NO READ exists