What stats do player prop tools actually use, and why those ones?

Most pages on this question list stats. This one is about the weighting, because the weighting is where the choices are and the choices are where these tools differ.

Here is what ours actually does.

The form score: 0.6 recent, 0.4 season

Our form number is:

0.6 x last 10 games + 0.4 x season

That is it. Where this season has no rate yet, last season’s stands in. No hidden term, no model output. The interesting part is why it is a blend rather than either half.

Season form alone is too slow. A player who changed something in July is still being described by April. By the time a season average moves enough to notice, the thing that moved it is months old.

A 10 game window alone is mostly noise. Ten games of a yes-or-no outcome is a very small number. “Six of his last ten” reads as 60% and is compatible with a true rate almost anywhere in the middle of the range. Treated as a fact it is one of the most misleading numbers in prop betting, and it is also the number most tools put in the largest font.

The split is a working compromise between those two failures, weighted toward the recent because that is the half carrying new information. We are not claiming 60/40 is optimal. We have not run an outcome study that would let us claim it, and a number arrived at by tuning against results we already saw would be worth less than the honest version.

The 20 game minimum, and why it bites unevenly

The picks our model selects automatically need at least 20 games of history. On the board, the L5, L10 and L20 rates appear once a player has 5, 10 and 20 games on file, and shorter histories are still shown.

That rule sounds mild. On one side of the board it is, and on the other it is severe:

Batters play nearly every day. Starters pitch every fifth. So the same rule excludes one batter in twenty and three pitchers in ten, and a pitcher who clears it clears it with a quarter of the history behind a typical batter’s number.

This is the most practical thing on this page. When you see a strikeout prop with a confident looking hit rate beside it, the sample behind that rate is usually a fraction of the sample behind a hits prop, and almost no tool tells you which you are looking at.

The trap that makes short-window stats look predictive

This one is worth more than the weighting, and we found it the expensive way.

Filtering a game log moves the window, not just the population. Take a player’s last 10 games and filter them for some condition, and you no longer have his last 10 anything. You have a stretch of games reaching further back in time, and everything that changed since then is now mixed into your result.

We tested this on one of our own ideas: a filter on the opposing starter, applied to unders. The filtered sample looked meaningfully different from the unfiltered one. Splitting it apart, 78% of the apparent effect was the window reaching further back, and 22% was the hypothesis we were actually testing.

Nearly four fifths of the finding was the calendar.

The fix is to hold the recent window fixed and vary only the thing you are testing. It sounds obvious written down. It is not what most “he hits 70% in this split” numbers have done, and it is the reason a split can look strong, be reproducible, and still be mostly an artefact.

What is deliberately not in the score

No rank of inputs by predictive power. We could tell you form matters more than matchup, or the reverse, and it would sound authoritative. We have not measured it, so we do not say it. This page describes how the number is built, not what predicts an outcome.

No claim about our results. Our public pick record started on 25 September 2026. Nothing here cites it.

No single metric explained twice. What a given stat means for a given market is a different page: home runs, total bases, strikeouts and RBIs each cover their own inputs. This page is the layer above them.

The short version

If you are evaluating any prop tool, ours included, three questions separate the useful ones from the rest:

1. What is the window, and is it weighted? A pure L10 number is noise with a decimal point on it.
2. What is the minimum sample, and what happens below it? Dropping thin samples is a harder choice than showing them, and it is the right one.
3. Does any split hold the time window fixed? If not, you are being shown a calendar effect with a story attached.

The Lab board shows the form number, the window behind it and the sample size on every prop, so you can apply those three questions to our own output rather than taking this page’s word for it.

If you want the narrower version of the first question, do L5 and L10 prop trends matter is about short windows specifically.

Form weighting is read from our own board code and the 20 game minimum from our model’s code. Sample depth figures measured 12 September 2026 across the players priced on that slate.