Every free-agency auction turns four or five lots into a fight, and the fight always feels like it means something — somebody else wants him, so he must be worth having. Matched against lots at the same place in the outside market’s own ranking, a contested lot does deliver more. It delivers about eleven season points, which is roughly a quarter of the ordinary spread between two players ranked the same, for a price a little over four and a half times as high.
A war lot changed hands six or more times on the way to closing; a quiet lot changed hands once or not at all. Comparing them raw is meaningless — wars happen to good players. So each lot is placed in a decile of MFL’s own average-auction-value rank, the outside market’s opinion formed before our auction started, and the arms are compared only inside a decile. Six deciles carry at least eight lots on both sides. This is the difference in season points, war minus quiet, within each.
Left column: where the AAV rank band starts — bars sorted by size of the gap
Averaged across the six deciles the gap is +10.9 season points, positive in five of them, with a bootstrap 95% interval of +0.9 to +21.5. So the crowd is right about the sign: the lot everybody wanted really was the better lot, over and above what the outside market already knew about him. The argument is entirely about size.
| AAV rank | War n | Quiet n | War pts | Quiet pts | Difference | War price | Quiet price | Multiple |
|---|---|---|---|---|---|---|---|---|
| 170–218 | 38 | 10 | 94.1 | 93.2 | +0.9 | 4.72M | 0.95M | 4.97× |
| 219–265 | 28 | 20 | 100.1 | 74.6 | +25.5 | 3.60M | 0.75M | 4.80× |
| 265–313 | 19 | 26 | 81.2 | 90.8 | −9.6 | 3.10M | 0.95M | 3.26× |
| 314–377 | 15 | 37 | 82.5 | 67.0 | +15.6 | 3.95M | 0.70M | 5.64× |
| 378–440 | 11 | 41 | 76.7 | 55.1 | +21.6 | 3.05M | 0.65M | 4.69× |
| 440–544 | 9 | 40 | 80.1 | 68.5 | +11.6 | 2.65M | 0.60M | 4.42× |
The number only means something against the spread it sits in. Two lots drawn from the same rank decile differ by a standard deviation of 42.4 season points, on a matched-cohort mean of 78.2.
So +10.9 is 0.26 of a standard deviation — about a seventh of a typical lot’s season, and comfortably inside the range that ordinary luck produces between two players the market ranked identically. It is the size of roughly one decile of ranking: on the open board, stepping from the 314–377 band up to 265–313 is worth +11.5 points. Winning the war buys you what moving up one rung of the market’s own ladder buys you.
It is at least a genuine shift rather than two or three lottery tickets. Trimming the top and bottom 5% of each arm leaves +8.8, and the gap is visible across the whole distribution — the war arm is ahead at every quantile, not just the ninetieth.
| Quantile | War lots | Quiet lots | Gap |
|---|---|---|---|
| 10th | −47.2 | −56.7 | +9.4 |
| 25th | −28.0 | −33.6 | +5.7 |
| 50th | +5.5 | −6.7 | +12.2 |
| 75th | +30.1 | +23.2 | +6.9 |
| 90th | +65.5 | +48.8 | +16.7 |
The clearest way to say it: inside a matched rank band, 29% of war lots finished under 60 season points against 40% of quiet ones. The fight lowers the bust rate by about eleven points of probability. It does not buy a different class of player.
Matched on rank, war lots cost 4.58× what quiet lots cost — a median of $3.51M against $0.77M — and return 1.15× the points. Divide one by the other and you get a headline: a war lot costs three to six times as much per point, in every single decile, with no reversals. That number is true and it is close to worthless.
It is worthless because the cheap arm wins per-dollar automatically. Points do not scale with price in any market with noise in it — a $0.60M lot that stumbles into 68 points will beat a $2.65M lot at 80 points on the ratio no matter which is the better roster decision. A per-dollar table measures the shape of the price curve, not the quality of a choice.
And the tempting corollary — four quiet lots outscore one war lot for the same money — is not a decision anyone can take. You cannot start four players in one slot. The auction hands out roster spots, and the question in front of a manager mid-lot is always what this spot should cost, never how many players a budget could theoretically buy. The quantity that survives both objections is the matched points difference, so that is the one this piece is built on.
Rank-matching answers “is the contested player better?”. It does not answer the question actually being asked mid-auction, which is “is this the right use of this money?” For that, hold the budget fixed and let the player vary.
Among lots that closed between $1.0M and $3.0M — the only band with a real sample on both sides — the 44 that were fought over returned 83.8 season points and the 38 that were not returned 83.9. The war lots were better players going in (median AAV rank 301 against 323.5). The extra ranking is exactly cancelled by what the fight charged for it.
Fix the player’s standing in the outside market and the fight adds +10.9 points at 4.58× the cost. Real, small, and expensive.
Fix the budget and the fight adds −0.1. Whatever the contested player has on the uncontested one at the same money, the premium has already taken back.
This second comparison is descriptive, not causal, and must not be read as one. Final price is downstream of the war — conditioning on it is conditioning on a consequence of the thing being measured, which is exactly the move that manufactures spurious findings elsewhere in this knowledge base. It cannot tell you what a war does. It can tell you what $2M has historically bought in each aisle, which is the shape of the actual decision.
Two seasons carry the entire result. Rank-adjusted, the war arm beat the quiet arm by +22.0 in 2021 and +22.9 in 2024, and by between −1.2 and +1.5 in the other four. The gap is 3 of 6 seasons positive, not six. An effect that lives in two years out of six is a good reason to hold the interval, +0.9 to +21.5, rather than the point estimate.
Half the cohort is one auction. 2020 supplies 441 of the 953 lots and 204 of the 329 wars — the inaugural auction, when every roster was empty and almost nothing went uncontested. Drop it and only two deciles still clear eight lots on both arms (+8.4); restrict to 2022–2025 and one decile survives, at +0.1. The design does not have the sample to stand without 2020.
The top of the market is unmeasured, and that is where wars actually hurt. Among lots ranked 1–100 there are 128 wars and zero quiet lots. There is no such thing as an uncontested stud, so this piece has nothing to say about whether to chase one. Every number here comes from ranks 170–544 — the middle of the board, where the walk-away decision is cheapest to get right and least costly to get wrong.
Part of the gap is roster tenure, not scoring. A player’s weekly points are recorded only while some franchise holds him, so a lot that busts and gets cut stops accumulating a record. War lots carry 0.72 more recorded weeks than matched quiet ones, and decomposing the +10.9 puts roughly a third of it in that extra tenure rather than in better weeks. On a per-week basis the difference is +0.76, with an interval of −0.05 to +1.59 that includes zero. An expensive player is kept; a cheap one is dropped. Some of what looks like production is patience.
The estimate moves with the matching rule. One-to-one nearest-neighbour matching on rank gives +6.5, +1.5 or +10.1 season points at calipers of 15, 25 and 40 ranks. Widening the war threshold to eight lead changes gives +5.9; narrowing quiet to zero lead changes gives +11.8. Nothing here reverses, but a finding whose size doubles on an arbitrary choice is a finding to state as a range.
WALK
Leaving a mid-board war costs about +10.9 points of expectation — one rung of the market’s ladder, a quarter of the noise between two identically ranked players. That is a real cost and it is a cost worth paying to keep four and a half times the money. Nothing in six seasons says the discipline is expensive.
NOT FREE
The other half of the instinct — that the war lot is overrated, that the room has lost its head — is wrong. The contested lot beat the uncontested one in five of six deciles and busted less often, 29% against 40%. Walking away is a price decision, not a read on the player.
The operational version: a war is information about the room’s appetite, not new information about the player. The outside ranking already knew most of what the fight is discovering, and the fight charges 4.58× to confirm it. Set the number before nomination, and treat lead changes as a signal to re-check that number, not to raise it.
A lot is one player put up in a season’s free-agency auction, taken from the auction ledger in the metrics store; 953 lots between 2020 and 2025 carry a final price, a lead-change count, an MFL average-auction-value rank for that season, and at least one week of scoring in this league’s own rules. Lead changes counts how many times the high bid moved to a different franchise, so it measures contest rather than duration or price. War is six or more, quiet is one or none; the 387 lots in between are excluded from the two arms but still set the decile boundaries. Points are the sum of a player’s weekly scores in the season he was bought, in BWAFF scoring, not a projection. Deciles are cut on AAV rank across the whole 953 so both arms face identical boundaries, and the six-decile average weights each decile equally. Intervals are bootstrapped over 20,000 resamples of both arms within each decile.
Four things it cannot see. Who else was bidding, and whether the war was against one determined rival or five casual ones — distinct leaders exists in the store and is not used here. What the money would otherwise have gone to, since a budget is a season-long constraint and this treats each lot alone. Contract term, which in this league converts a price into a multi-year commitment and is the largest missing variable by some distance. And any lot from 2026, deliberately excluded — the season has not been played, so its points do not exist yet.
Confidence ceiling 0.45 — realised, not forecast. This describes what 953 historical lots did. Two of six seasons carry the effect, half the sample is one unrepeatable inaugural auction, and the matched estimate moves by a factor of two under reasonable changes to the matching rule. Treat +10.9 as the centre of a wide range, not as a coefficient.