Reading Projection Bands: When P10-P90 Overlap Turns a Start/Sit Call Into a Lean

A start/sit decision usually gets made on one number. Two players, two medians, the higher one starts, and the manager moves on to the next slot.

The median is the smallest part of what the weekly model publishes. Around it sits a P10-P90 band, and that band is the model saying how far players at this projection level have historically finished from where they were projected. When two bands overlap through most of their length, the gap between the medians is smaller than the week-to-week noise at that position.

That situation is a lean. The higher median is still the better guess, and it is still the right start most of the time. What changes is how much a manager should be willing to pay to act on it — how much waiver capital, how much roster churn, how much regret when it goes the other way.

What the Band Actually Measures

The band's shape comes from history, and it comes from a bracket. At each position, a player's own weekly projection drops him into one of three brackets, and every player in a bracket prints the same proportional band, fitted per scoring format. A projection near the top of the position gets the tight band that top-bracket players have historically earned. A projection near the bottom gets the wide one.

This is the part most managers read backwards. A wide band on a given player is not the model flagging his depth chart, his target competition, his injury history, or a coaching staff that keeps changing its mind. The weekly model has no input for snaps, routes, usage shares, depth charts, practice reports, or opponent defense. Width is a property of the projection level, and a player's own week moves his band only by moving his projection across a bracket line.

The practical consequence: two players with similar projections carry similar band widths by construction. Overlap between them is the normal state at the back of a lineup, not an anomaly worth investigating. What's worth investigating is the rare pair whose bands barely touch.

One more mechanic that changes how the P50 reads. The weekly model blends this season's scoring — recent weeks weighted up — with the last two seasons and a position mean, then scales the result by the game's implied team total, and it publishes a median. The middle outcome, not an average. Half the time the player finishes under it, which is exactly the property that makes a narrow gap between two medians so weak a signal on its own.

Running the Arithmetic on an Invented Week

Take an invented stretch of a week's receiver projections — numbers made up to show the shape, not this week's board and not any real player. Four flex candidates, each an archetype:

The every-week starter projects 14.8 with a band of 8.9 to 22.2. The second option, a WR2 on a team with a high implied total, projects 12.6 with a band of 5.7 to 21.4. The third, a volume receiver on a bad offense, projects 11.9 with a band of 5.4 to 20.2. The fourth, a rotational piece the room streams, projects 8.4 with a band of 3.8 to 14.3.

Check the shapes against the bracket rule before trusting anything else. The bottom three share one bracket, so they share one proportional band: each P10 sits about 55 percent below its median and each P90 about 70 percent above it. The starter's projection puts him in the tighter bracket — P10 40 percent under the median, P90 50 percent over. That is all the band widths say. None of them is a statement about a depth chart.

Now do the comparison the model can actually support. Option two against option three is a 0.7-point median gap between bands that overlap from 5.7 to 20.2 — 92 percent of option two's band and 98 percent of option three's. Option two's P10 of 5.7 sits far below option three's median of 11.9. That is a lean: starting the 12.6 is still the better call, and it will lose to the 11.9 nearly as often as it beats it.

Option one against option four is a different object. The medians sit 6.4 points apart, and although the bands still cross between 8.9 and 14.3, the starter's P10 of 8.9 clears option four's median of 8.4 — the bad version of the starter still beats the typical version of the streamer. That is a start, and a matchup read alone should not move it.

The rough cutoff that falls out of this: when the higher player's P10 sits at or above the lower player's P50, the call is a start. When the higher player's P10 sits below the lower player's P50 — when the bad version of the better option is worse than the typical version of the worse one — the call is a lean. That threshold is judgment fitted to how the brackets space out, not a backtested rule, and it should be read as one.

What a Lean Changes About Your Week

A lean and a start get played the same way in the lineup. They get played differently everywhere else.

On a lean, stop spending. Do not burn a waiver claim to improve a slot where the model is telling you the improvement is inside its own error bars, and do not trade a bench piece to resolve it. That claim is worth more pointed at a slot where the gap actually clears the bands, or at a role change the model has no input for. The reallocation argument gets its own treatment on the bench allocation page.

On a lean, your own information is worth something. The weekly model drops players whose current designation is Doubtful, Out, IR, PUP, or suspended, and leaves a Questionable player projected and unchanged. It does not read a beat report about a snap count, and it does not price a route-running role that changed three weeks ago. A manager who has that read is adding the only thing the model isn't carrying, and a 0.7-point median gap is exactly the situation where that addition should decide it.

On a start — bands mostly clear, the better player's P10 above the other's median — your read is worth less than you think. Overriding a separated pair on a matchup feeling is how managers lose points slowly across a season.

Band width also travels with projection level, which means the lean condition is structurally more common at the bottom of a lineup than the top. Flex slots, streaming tight ends, and second-half-of-the-bench decisions live in the wide brackets. Two RB1-level medians separated by two points will often clear; two flex-level medians separated by two points almost never will. The mechanics of that variance spread get worked through on the variance players page.

Where to Read the Numbers

Everything above is mechanism. The current medians and bands live on the start/sit surface, which reruns daily off the weekly model — a Sunday-morning designation change moves the page.

The grading works on a different clock, and the distinction matters when you go check our record. The accuracy ledger grades each week's first posting, frozen before that week's first kickoff. A later rerun can change what the page shows you and never changes what the week was graded on. The ledger covers the weekly model only; the season model behind the draft board has its own backtest on the methodology page, including the losses — the v2 receiver backtest beat a prior-season baseline by 0.0048 Spearman, a margin thin enough that calling it a win requires saying how thin.

That honesty runs in both directions here. A band is a minimum on risk, not a full accounting of it — the real distribution of outcomes is at least this wide, and the model's blind spots at the bottom of the roster are the reason a lean is a lean.

This piece covers one slot at a time. The season-long version of the same idea — which players carry structurally wide outcomes before you ever set a lineup — is drafted, not managed weekly. Start on this week's start/sit page for the live medians, then read the public accuracy ledger and decide from the record how hard to lean on any of it, this piece included.