How our Week 1 rankings did

Starter picks

A starter call is a player we ranked inside the starters a 12-team league fields at each position, not counting flex spots: 12 quarterbacks, 24 running backs, 24 wide receivers, 12 tight ends. It is a hit when he finished inside the same line at his position. A player we ranked as a starter who did not play counts as a miss, because the board's rankings already account for the chance a player sits.

36 of 72 players we ranked as starters finished as starters (50%). The outside projection: 35 of 72 (49%).

Of the 72 players who actually finished as starters, we had ranked 36 inside the line (50%). That is the same count read the other way: we rank exactly as many starters as the league fields, and exactly that many finish there unless players tie.

1 player we ranked as a starter recorded no statistic and counts as a miss. Brock Bowers was ruled out in a status capture taken after the board went up, which a board published before the week cannot see: check the current board before lineups lock.

Hit rate: of the players we ranked as starters, the share who finished as starters. Capture rate: of the players who finished as starters, the share we had ranked as starters, including players we did not rank at all.

Start/sit calls

Both rankings are graded on the same pairs. For each position, the pool is every player either ranking put near the starter line (QB and TE ranked 7 to 18 and RB and WR ranked 13 to 36), and every pair in it is compared once: a call is right when the player ranked higher scored more. A player who did not play scored zero, and a player a ranking did not list counts as ranked below everyone it did.

LMD put the higher scorer first in 519 of 991 scored pairs (52%). The outside projection: 535 of 991 (54%). Picking at random would be right half the time.

997 pairs in all. Left out for every ranking: 3 pairs whose players scored exactly the same, 3 pairs the outside projection ranked equally and 0 pairs LMD ranked equally.

Rank misses

How many places each player finished from where we ranked him, for the players we ranked as starters who recorded a stat line. A receiver we ranked WR8 who finished WR14 is 6 places off.

Across 71 players, the average miss was 16.3 places and the median 10. The outside projection: 16.4 and 11 over its own 71 starters. A few large misses pull the average up; the median is the typical player.

A player with no stat line has no finish, so he is left out here and counted as a miss in the starter calls above.

The same measure over the players ranked inside the top 24 at their position who recorded a stat line, as earlier reviews reported it. A receiver ranked WR8 who finished WR14 is a miss of 6 places.

Rank correlation is Spearman's: 1 is the right order, 0 is no relationship. Inside the top 24 it is a hard test, because those players are close together; over the whole board it is an easier one, because it includes backups who were never near a start. It is not pooled across positions, which would correlate four different orderings.

Top 3, 12 and 24 picks

The same count at every finish line the board prints: of the players we ranked inside the top 3, 12 or 24 at their position, how many finished there, all four positions together. The chances printed beside each rank are graded as probabilities further down.

Top 24 is a wide net at quarterback and tight end, where a 12-team league starts 12, so the starter calls above are the better read there.

Every hit and miss list

Probability detail

Beside each rank the board prints a chance of finishing in the top 3, 12 and 24 at his position. If those chances are right, a group of players each given 60% should get there about 60% of the time, and the total number of finishers should land near the sum of the chances. They are graded as probabilities, not as yes-or-no calls at 50%.

Typical range is the expected count plus or minus two standard deviations of what the chances imply. Brier is the mean squared gap between each chance and what happened; lower is better. "Same % for all" gives every player at a position the same chance, and "rank as yes/no" says 100% inside the threshold and 0% outside: two simple forecasts the chances should beat. Groups under 10 players are shown faded: too few to read.

• The number of Top 3, Top 12 and Top 24 finishes landed inside the range the percentages implied (12 against 11.9, 47 against 45.7 and 93 against 88.7).
• At every threshold the percentages scored better than reading the rank as a yes-or-no call, so the odds columns said something the rank alone did not.
• One position at a time, the percentages lost to a simple forecast at Top 3 WR (0.025 against 0.024 for the same chance for all) and Top 12 TE (0.201 against 0.185 for the same chance for all). The pooled figures above do not show it.
• No group of 20 or more players finished outside the range chance allows for the percentages it was given.
• Of the 51 players given a 30-50% chance of a Top 24 finish, 47% got there. A percentage on the board is a rate across many players like him, not a call on one.
• 142 of the 400 ranked players recorded no statistic. Their chances are part of every percentage above, which is why the whole board is graded rather than only the players who played.
• Of the Top 24 finishes the board expected, players who recorded no statistic were given 6.3 of 24 at quarterback (67 players) and 2.5 of 23.2 at running back (55 players). Chances spread onto players who did not take the field are chances the starters' percentages did not get, so read a small percentage beside a backup's name with that in mind.

Were firmer calls more reliable?

No confidence rating was published with the Week 1 board, so there is none to grade. What the board did publish is how one-sided each player's chance was. Here the players ranked in the top 24 at each position are grouped by their Top 12 chance: the further from an even split, the firmer the call. The median rank miss counts only those who recorded a stat line.

In every group large enough to read, the share landing on the favoured side was within the range its percentages implied. Firmer and closer calls missed by the same median of 10 places, which in one week is not evidence either way.

Other scoring formats

How this review was graded

The benchmark. The benchmark is an outside projection service LMD's board uses as one input, captured 32 seconds before this board and scored the same way. The board gives it at most 35% of its weight. It is the newest outside projection captured at or before this board, and never one captured after it. Its projected stat lines are priced with our scoring table, ranked within each position, and scored by exactly the calculations above against the same finishes. It is not a consensus of rankers, and its vendor is recorded in the audit file committed with this review.

The board. 3 boards were archived for Week 1. This is the last one generated before the first kickoff (September 10, 00:20 UTC), 5.7 hours ahead of it, and it is the board readers saw going into the week. The September 13 refresh was generated after games had started and is not graded here.

The population. All 400 players the board ranked at QB, RB, WR and TE. A finish is counted over every player in the league who recorded a statistic (360 stat lines, including 102 players the board did not rank), priced with the same Full PPR scoring the board projected. Players who tie share the better place.

Starter calls. The starter line is Top 12 at quarterback and tight end and Top 24 at running back and receiver. The hit rate counts the players ranked inside it who finished inside it; the capture rate counts the players who finished inside it, ranked by us or not, whom we had ranked inside it. Top 3, 12 and 24 use the same arithmetic.

Start/sit calls. One set of pairs for both rankings. The pool is every player either ranking put inside half a starter line either side of it (QB and TE ranked 7 to 18 and RB and WR ranked 13 to 36) -- the union, so each is graded on the players the other nominated too -- and every pair in it is formed once. A pair is scored for both rankings or for neither: only when the two players' points differ and each ranking puts them in a strict order. A player a ranking did not list counts as ranked below everyone it did. The band is set by the starter line, not chosen after the games.

Rank error. The number of places between a player's published positional rank and his finish, over the players ranked inside the starter line who recorded a stat line. Earlier reviews reported it over the top 24 at every position, and that table is kept under the rank error section.

Players who did not play. 142 ranked players recorded no statistic. 24 were ruled out in a status capture taken after the board was published (September 13, 18:12 UTC) and 118 had no ruling; 53 of those were quarterbacks. A player we ranked as a starter who did not play is a missed starter call, and scores zero in a start/sit comparison. Players who did not play are left out of the rank error, which needs a finish, and counted as misses in the probability checks, because the board's rankings and percentages already include the chance a player sits.

What is not adjusted. Nothing is corrected for in-game injuries, game script or garbage time: a player who left early is graded on what he scored. Player pages chart the last board before each player's own kickoff, so for later games their Week 1 projection can come from a newer board than this one.

Sample size. This is one week. A few players' afternoons move every figure on this page, so read it as one data point rather than a verdict, and nothing here has been used to retune the model.

Sources and evidence

• Model rankings-v1.3.0, configuration d6d390f3d5c15fe3, board of 2026-09-09 generated September 9, 18:40 UTC.
• Archived board (Full PPR): rankings/2026/week-01/artifacts/rankings-v1.3.0/2026-09-09/weekly-ppr.json, SHA-256 a00a60fc443cb4f4307d1f9c05e6a1df53865677a8806f8d2a69dce93fdd4442.
• Results: nflverse weekly player statistics, published September 15, 14:21 UTC (CC BY 4.0). 4 stat lines could not be matched to our player catalog and still count toward finishes.
• Rulings: balldontlie status capture of September 13, 18:12 UTC, used only to label players who did not play.
• Benchmark: outside projection captured September 9, 18:39 UTC, 414 rows (3 not matched to our catalog, none inside a graded depth), SHA-256 80dabf5a0f82c63462b65e0429249840e54600f73595214f046aeefe254c70cc.
• Per-player evaluation SHA-256 3fb030be74511e25f5cfb0d6e3238cbd2c988c0d180f3b2aca4d97fbc4b909a8.