Original research / September 2026 / 12 min read
The number the market hung
How often an NFL middle actually landed, measured against the closing price that created it, across 6,719 regular-season games.
Written and reviewed by Alexander Korsager · last reviewed September 11, 2026 · corrections
The finding
4.80%
how often the tightest spread middle landed over 26 seasons, against the 4.76% it has to clear to break even
- Landed, over 26 seasons
- 4.80%
- Needed to break even at −110 on both legs
- 4.76%
Clears it by 0.04 of a percentage point, across 3,751 games.
Both bars are drawn on the same scale, running from zero to 4.80%.
The numbers behind it
- 6,719
- NFL regular-season games, 2000 to 2025, each joined to its closing spread, closing total and final score
- 0.0% to 10.6%
- how often the same one-point spread middle landed, across the 13 whole numbers the market hung at least 100 times
- 2.0 times
- what the published key-number table reports for a one-point middle on Away −3, against what it did
What we found
The findings, in 5 sentences
- Across 6,719 NFL regular-season games from 2000 to 2025, a one-point spread middle bought on a whole-number closing line landed 4.80% of the time. Break-even on two legs priced at −110 is 4.76%.
- Knowing the closing spread beats ignoring it entirely by about a fifth: the tightest middle landed 4.80% against 4.02% for the same results ignoring the line. Most of that is knowing which side is favored; the exact number on its own is worth about 2.6% more in relative terms once the side is already priced in (4.68% to 4.80%). Knowing the closing total is worth 0.06 of a percentage point, which is nothing.
- The published key-number tables everyone quotes count a margin in both directions. With the market closing at Home −3, the home team won by exactly three in 10.59% of 614 games, against a 14.99% two-way figure. At Away −3 the same table overstates by 2.0 times.
- Which number you middle matters far more than how wide you make it. A close of Home −3 landed a one-point middle 10.59% of the time; a close of Home −4 landed one 0.99% of the time.
- The average game finished 0.66 points above the closing total, and overs still went 3,276 and 3,345, a losing 49.5% against a 52.4% break-even. Both are true at once, and the second is the one that decides money.
For journalists and researchers
Every figure on this page is read from a published result set rather than typed, and the data behind it is yours to download. Republish what you find, including commercially, with credit.
- Download the datasetCSV, 102 rows
- How this was measuredThe method, the sample, what it cannot show
- Sources and how to cite itReleased under Creative Commons Attribution 4.0 International
Found an error, or need a figure this page does not print? We would rather hear about it than not: how we handle corrections.
A middle is the most attractive-looking bet in sports betting. Take the favorite at one number with one book, take the underdog at a longer number with another, and a band of results collects both tickets. Miss the band and you win one bet and lose the other, which costs you only the juice on the loser. Small downside, double upside: the shape of the thing sells itself.
What no free tool in this market publishes is the number that decides whether it is worth doing. Every middle calculator we have found, ours included until now, takes the window and the prices and returns a payout, then asks the player to supply the hit rate from their own head. The hit rate is the entire question.
So we took it out of the reader's head. This study joins the closing spread and closing total of every NFL regular-season game from 2000 through 2025 to the final score of that game, 6,719 games, and counts how often the middle window actually landed, given the market number that created it. That last clause is the part we could not find published anywhere. Key-number frequency tables are everywhere, and betting guides do print rough hit-rate ranges for a middle of a given width. What is absent is the same question asked one line at a time, of the price that was actually on offer, with the side named and the count printed.
What a middle is, and what this counts
Buy the total Over at 43.5 with one book and Under at 45.5 with another. A game finishing 44 or 45 collects twice. Anything else collects once and pays the losing leg's juice. The gap between the two numbers, two points here, is both the width of the middle and the count of results that win it.
Both numbers are always half-points in this study, which means neither leg can push and no tie needs a rule. That is a modeling choice and it is stated here rather than buried, because it is the choice that makes every figure below a clean count rather than a count with an asterisk.
Where the window sits is decided by the market, not by the bettor, and this turns out to matter more than it sounds:
- A whole closing number (a 3, a 44) can be middled symmetrically only at odd widths. One book half a point each side of a 3 gives you a one-point middle that wins on a margin of exactly three.
- A half-point closing number (a 3.5, a 44.5) can be middled symmetrically only at even widths. The tightest available is two points wide, and it needs the two books to be two full points apart rather than one.
So a whole-number close is not merely a different price from a half-point close, it is a different product. It is where the cheapest middle in football is on sale, because a single point of disagreement between two books puts the whole window on one number. A half-point close cannot be middled that tightly by anyone, at any price. Nothing on this page averages the two together.
Every figure below is a count over the same sample: 6,719 games, each carrying a closing spread, a closing total and a final score. No model, no simulation, no estimate.
Finding 1
The closing number tells you a lot about a spread and almost nothing about a total
Here is the comparison the study was built to make. For every game, take the middle window sitting on that game's own closing number, and ask two questions of it. How often did the result land inside it, given that line? And how often did those same result values occur across all 6,719 games, ignoring the line entirely? The second is the reading a player gets from a published key-number table, including the one on our own middle calculator.
| Market and middle | Given the line | Ignoring the line | Games | Landed |
|---|---|---|---|---|
| Spread, 1 pt, whole | 4.80% | 4.02% | 3,751 | 180 |
| Spread, 3 pt, whole | 11.20% | 8.91% | 3,751 | 420 |
| Spread, 2 pt, half | 8.22% | 6.85% | 2,968 | 244 |
| Spread, 4 pt, half | 14.79% | 11.86% | 2,968 | 439 |
| Total, 1 pt, whole | 2.88% | 2.82% | 3,407 | 98 |
| Total, 3 pt, whole | 9.30% | 8.19% | 3,407 | 317 |
| Total, 2 pt, half | 5.68% | 5.50% | 3,312 | 188 |
| Total, 4 pt, half | 12.05% | 10.84% | 3,312 | 399 |
The shaded band behind each bar is its 95% Wilson interval on the given-the-line rate.
Each row is a market, a middle width and the parity of the closing number it sits on, and the tick on each bar is that row's rate ignoring the line. Whole closing numbers take odd middle widths and half-point ones take even widths, because only that pairing puts the window symmetrically on the line. The two are different products and are never averaged together here. The band behind each bar is what the sample size alone can support: on the four total rows it sits close to the ignoring-the-line tick, which is the more precise version of the finding above that conditioning on a total buys next to nothing.
Source: actfap.com, original analysis
For spreads, conditioning on the line beats ignoring it entirely by roughly a fifth. The tightest whole-number middle landed 4.80% of the time against 4.02% unconditionally, and the same gap holds at every width and both parities: 8.22% against 6.85% at two points, 11.20% against 8.91% at three.
That fifth bundles two things a bettor could know, and they are not the same size. A spread's closing number tells you which side is favored and by how much, and most of the fifth is the first of those. Once the side is already priced in, using a baseline built only from games on that same side rather than the whole sample, the exact number on its own is worth about 2.6% more in relative terms (4.68% to 4.80%), not roughly a fifth. That is real and worth having, but it is a different, much smaller claim than "the closing number is worth a fifth," and this page previously did not separate the two.
For totals it is worth nothing either way, since a total has no side to know. The tightest whole-number middle landed 2.88% of the time against 2.82% unconditionally, a gap of 0.06 of a percentage point over 26 seasons. Of the 27 closing totals the market hung often enough to report, 13 came in above their unconditional rate and 14 came in below. That is a coin flip. For spreads the same count is 16 of 24.
This is worth being precise about, because we published the opposite assumption ourselves. The middle calculator's own note says a game the market hung at 44.5 lands on 45 or 46 more often than an average game does, so an unconditional figure reads low for a window sitting on the market's line. The direction of that is right. The size of it, measured, is about 7.3% of what the same reasoning is worth on a spread.
Why the two markets differ
The mechanism is visible in the dispersion. Knowing the closing spread removes 18.6% of the variance in the final margin, taking the standard deviation from 14.6 points to 13.2. Knowing the closing total removes 10.3%, from 14.1 points to 13.4.
Neither is much. An NFL result is enormously noisy around the number the market hangs on it, and a middle is a bet that the noise lands in a band a point or two wide. But the spread has something the total does not: football margins pile up on three and seven, and the market hangs spreads on three and seven precisely because they do. Conditioning on the line and conditioning on a key number are close to the same act. Totals have no comparable structure, so the market's number buys you very little beyond knowing roughly where the distribution sits, which you already knew.
Finding 2
The table everyone quotes is a ceiling, and it is not a tight one
Published margin tables, ours included, count a game decided by three points whether the home team or the road team won by it. That is the right way to publish a margin distribution and the wrong number for a spread middle, because a spread middle names a side. If you bought Home −2.5 and Away +3.5, a road team winning by three loses both your bets.
So the familiar figure is a ceiling. What it is a ceiling on, we can now say. With the market closing at Home −3, the home team won by exactly three in 65 of 614 games, 10.59%. The two-way table reports 14.99%.
The ceiling is loosest when the road team is the favorite.
| Closing spread | Wins on | Games | Landed | Measured | Two-way table says | Overstated by |
|---|---|---|---|---|---|---|
| Home −3 | home by 3 | 614 | 65 | 10.59% | 14.99% | 1.4 times |
| Away −3 | away by 3 | 464 | 35 | 7.54% | 14.99% | 2.0 times |
| Home −7 | home by 7 | 295 | 19 | 6.44% | 9.09% | 1.4 times |
| Away −7 | away by 7 | 152 | 6 | 3.95% | 9.09% | 2.3 times |
| Home −6 | home by 6 | 213 | 11 | 5.16% | 6.01% | 1.2 times |
| Home −4 | home by 4 | 202 | 2 | 0.99% | 4.91% | 5.0 times |
A published margin table counts a game decided by three points whether the home team or the road team won by it. A spread middle needs one named side, so that table can only put a ceiling on the hit rate. The gap between the two columns is how loose the ceiling is.
The gap is not constant, and the pattern in it is the useful part: the ceiling is loosest when the road team is the favorite. At Home −3 the published figure overstates by 1.4 times; at Away −3 it overstates by 2.0 times. At seven the same split runs 1.4 times against 2.3 times.
Home-field advantage is why. A road team favored by three is a strong team held down by the venue, and its margin distribution is pushed left relative to a home team laying the same three. It lands on the number less often, and the two-way table, which cannot see the difference, reports the same figure for both.
A player pricing a middle off the published table is therefore not slightly optimistic. On a road favorite at three they are reading a number 2.0 times the real one, which is the difference between a bet that clears the juice and one that does not.
Finding 3
Which number you middle matters far more than how wide you make it
The instinct with a middle is to widen the window. The data says the first question is not how wide, it is which number.
Widening a middle on a dead number by a full point does not recover that; moving to a live number does.
| Closing spread | Hit rate | Games | Landed |
|---|---|---|---|
| Home −3 | 10.59% | 614 | 65 |
| Away −3 | 7.54% | 464 | 35 |
| Home −7 | 6.44% | 295 | 19 |
| Home −1 | 0.73% | 273 | 2 |
| Away −1 | 1.99% | 251 | 5 |
| Home −6 | 5.16% | 213 | 11 |
| Home −4 | 0.99% | 202 | 2 |
| Away −7 | 3.95% | 152 | 6 |
| Home −10 | 7.95% | 151 | 12 |
| Away −6 | 1.79% | 112 | 2 |
| Away −4 | 5.41% | 111 | 6 |
| Home −5 | 0.00% | 108 | 0 |
| Home −9 | 0.00% | 107 | 0 |
The upright line on each bar is break-even at −110 on both legs, 4.76%.
The shaded band behind each bar is its 95% Wilson interval on the hit rate.
Each middle here is bought half a point either side of the closing number. Bars reaching past the line cleared the juice over these 26 seasons; bars short of it did not. Most of these samples are too thin to say more than that: only two bands, Home −3 and Away −3, sit entirely above the line, and only the first clears it by a margin wide enough not to be noise itself. On several other rows the band spans both sides of the line, which is the honest reading a bare bar cannot give.
Source: actfap.com, original analysis
The top and bottom of that table are not close. A one-point middle on Home −3 landed 10.59% of the time. The same middle on Home −4 landed 0.99% of the time, 2 of 202 games, and on Home −1 it landed 0.73% of the time. Widening a middle on a dead number by a full point does not recover that; moving to a live number does.
Two whole numbers the market hung often enough to report never landed a one-point middle at all in 26 seasons: Home −5 (0 in 108) and Home −9 (0 in 107). Neither is as remarkable as it looks: given each line's own baseline rate, the chance of seeing zero by chance alone is 13.8% at Home −5 and 37.7% at Home −9, so this reads as a small sample rather than a structural fact about where NFL games finish.
The reason is that NFL margins are not smooth. Three and seven are where a field goal and a touchdown put games, and four, five and nine are the gaps between them. The market hangs lines on all of these because it is pricing the teams, not the shape of the margin distribution, so a player can end up buying a window over a number the sport almost never produces.
What the circulating ranges hide
Betting guides do publish hit rates for middles, and they publish them by width. A widely repeated version gives roughly 3% to 5% for a one-point middle, 8% to 12% for two points and 12% to 18% for three, disclaimed as approximate and carrying no sample, no source and no method.
Those ranges are not wrong on average. Our pooled one-point figure, 4.80%, sits inside the first of them. What the range cannot express is that the lines underneath it run from 0.00% on Home −5 to 10.59% on Home −3, across the 13 whole numbers the market hangs often enough to measure. The bottom of that list is zero, not a small number.
A hit rate quoted without the number it applies to is an average over products that behave nothing alike. That is the practical case for doing this one line at a time rather than one width at a time, and it is the reason this page's tables have a row per closing number rather than a row per width.
Total middles, by the number
Totals behave differently again, and more tamely, which is the same finding as above seen from the other side. They have no dead numbers of the kind spreads do, and no key numbers either. The most-hung half-point close, 43.5, landed a two-point middle 6.27% of the time against 7.34% unconditionally. The most-hung whole number, 44, landed a one-point middle 4.18% of the time against 3.84%.
| Closing total | Bought at | Games | Landed | Given the line | Ignoring the line |
|---|---|---|---|---|---|
| 43.5 | Over 42.5 / Under 44.5 | 319 | 20 | 6.27% | 7.34% |
| 44.5 | Over 43.5 / Under 45.5 | 290 | 19 | 6.55% | 6.37% |
| 41.5 | Over 40.5 / Under 42.5 | 260 | 17 | 6.54% | 5.28% |
| 46.5 | Over 45.5 / Under 47.5 | 233 | 11 | 4.72% | 5.19% |
| 45.5 | Over 44.5 / Under 46.5 | 229 | 12 | 5.24% | 4.54% |
| 42.5 | Over 41.5 / Under 43.5 | 213 | 14 | 6.57% | 4.99% |
| 47.5 | Over 46.5 / Under 48.5 | 212 | 12 | 5.66% | 6.03% |
| 40.5 | Over 39.5 / Under 41.5 | 194 | 15 | 7.73% | 7.13% |
| 48.5 | Over 47.5 / Under 49.5 | 180 | 8 | 4.44% | 4.87% |
| 37.5 | Over 36.5 / Under 38.5 | 175 | 7 | 4.00% | 5.70% |
| 39.5 | Over 38.5 / Under 40.5 | 167 | 7 | 4.19% | 5.36% |
| 38.5 | Over 37.5 / Under 39.5 | 135 | 4 | 2.96% | 4.06% |
| 49.5 | Over 48.5 / Under 50.5 | 125 | 7 | 5.60% | 4.26% |
| 36.5 | Over 35.5 / Under 37.5 | 115 | 6 | 5.22% | 6.04% |
The last two columns are the whole finding for totals. They are close to each other on most rows, and on about half the rows the market's own number came in below the rate you would have got by ignoring it.
Finding 4
What it costs to be wrong
The modal outcome of a middle is that it does not land. At the tightest whole-number spread window, the most favourable cell in this entire study, that happened 95.20% of the time. At the tightest total window it happened 97.12% of the time. A middle is not a bet that usually works and occasionally does not. It is a bet that usually does not work, priced so that not working is cheap.
How cheap decides everything. One unit on each leg at −110: a middle collects both legs, a miss collects one and loses one, for a net loss of about a tenth of a unit. Break-even is 4.76%, or exactly one in twenty-one. That is the bar every figure on this page should be read against, and it moves fast: at −120 on both legs the same bet needs 9.09% and almost nothing here clears it.
A window that landed 8% of the time is worthless if the only way to buy it was at a price needing 10%.
- Spread, whole-number close, one point wide: 4.80% against a 4.76% break-even. Historically a coin flip against the juice, not an edge.
- Total, whole-number close, one point wide: 2.88%. Well under break-even, over 26 seasons and 3,407 games.
- Spread, half-point close, two points wide: 8.22%, comfortably above break-even, but it requires two books a full two points apart, which is a rarer and usually worse-priced thing to find than the one-point disagreement above it.
Those three lines are the honest summary of the study, and none of them is a strategy. The frequencies here are the numerator. Whether any of this clears a real price depends on the two prices you were actually offered, and this study does not observe prices offered, only the closing consensus. A window that landed 8% of the time is worthless if the only way to buy it was at a price needing 10%.
So should you just bet overs, or favorites?
A reader will notice that the average game in this sample finished 0.66 points above its closing total, and reasonably ask whether that is a signal. It is not, and the reason is a good lesson in reading distributions.
Overs went 3,276 and 3,345 with 98 pushes, a 49.48% strike rate. That is a losing side outright, before the juice, and the juice needs 52.38%. The average is above the line and the median game is below it, because a handful of very high-scoring games pull the mean up while most games finish just under. Favorites against the spread went 3,188 and 3,325 with 206 pushes, 48.95%, also losing.
Over 26 seasons the closing NFL line is close enough to right that no naive side of it was profitable, and that is the context every number on this page sits in. Nothing here identifies a bet worth making. It identifies how often a thing happens, which is the input people were previously supplying from intuition.
One more thing worth saying plainly
Sports betting loses money for most people who do it, over time, by design. Nothing in this study changes that and nothing in it should be read as encouragement. A middle in particular looks like a structure that beats the maths, and the maths on this page says it mostly does not.
This site carries no affiliate links and takes no operator money, which is the only reason we can publish a finding this unflattering to a popular strategy without it costing us anything. Betting is for adults only and is legal in some US states and not others. Free, confidential help is linked below if it has stopped being entertainment.
How we know
Methodology
Written to be read by someone trying to find a hole in it. Where a limit cuts against the study, it is stated here rather than left for a reader to discover.
Sample and source
Every NFL regular-season game from 2000 through 2025 that was played and carries both a closing spread and a closing total: 6,719 games. No game in that window was dropped for a missing price, so there is no imputation anywhere in this study and no exclusion rule doing hidden work.
The closing prices and final scores come from one file, games.csv in the nflverse project's nfldata repository, retrieved 2026-09-11. It is a plain public file. Fetching it is a single request for a URL the project publishes for reuse: nothing here reads a page behind an access control, a paywall or a robots disallow.
The field definitions matter enough to quote. nflfastR's reference manual defines spread_line as the closing spread line for the game, positive when the home team was favored, sourced from Pro-Football-Reference, and total_line as the closing total line for the game, same source. The data repository's own documentation describes the same two fields without the word closing and without the provenance, so the manual is what this study cites.
The file is pinned by hash. The exact snapshot behind these figures is SHA-256 52167cd5edf6af92…, recorded in the study's working files, so a later run can tell an upstream change apart from a mistake of ours.
A source we checked and did not use
A well-known free archive of historical NFL odds publishes both opening and closing prices, which would have made an attractive second source. It returns HTTP 403 to a direct request, and its robots file reserves its content against automated reuse. A 403 is an access control, and our research rule has no usefulness exception, so it was not used and not worked around. Recorded here so the absence reads as a decision rather than an oversight.
How a middle window is defined
A middle of width d is bought at two half-point numbers d points apart, placed symmetrically on the closing number. Its window is the d consecutive results strictly between them, and because both bet numbers are half-points, neither leg can push.
Symmetric placement exists for only one parity of d per line, as described above: odd widths on whole closing numbers, even widths on half-point ones. The study reports widths 1 and 3 on whole numbers and 2 and 4 on half-point numbers, and never pools across the two, because a one-point middle and a two-point middle are different bets requiring different amounts of disagreement between books.
Spread results are home-relative throughout. A margin of +3 is the home team winning by three and −3 is the road team winning by three, and they are counted as different outcomes, which is the whole of Finding 2.
Which closing numbers are reported, and which are not
A closing number appears in the per-line tables only if the market hung it at least 100 times in the sample. Below that the standard error swamps the estimate: a line seen forty times can put a confident-looking hit rate on a page off two games.
That threshold reports 24 of 78 distinct closing spreads, covering 5,231 of 6,719 games, and 27 of 64 distinct closing totals, covering 5,643. The pooled figures in Finding 1 use every game, thin lines included, so the two readings are not the same denominator and are not presented as though they were.
Even at 100 games a cell is not precise. A rate like Home −6's one-point middle, 5.2% on 213 games, carries a 95% interval running from 2.9% to 9.0%, roughly 6 points wide. The per-line tables print the raw count alongside every rate for exactly this reason: read 2 in 202, not 0.99%.
What we checked this against, including where it disagrees
The sample was validated against an entirely separate compilation of the same 26 regular seasons: the published margin and total tables already transcribed on our middle calculator, which come from a third party's own count. Recomputing both distributions from this sample's game scores reproduces them across roughly 110 cells.
It does not reproduce them exactly, and the difference is reported here rather than smoothed. Two cells differ, each by one game: margin 4 (331 published, 330 computed) and total 50 (151 published, 150 computed). That is one game, with a four-point margin and a fifty-point total, present in their count and absent from ours. Our source carries 271 rows for the 2022 regular season rather than 272, because the abandoned Bills at Bengals game is absent from it, but that game stood 7-3 when it was called, so it is not the row in question and we could not reconcile the difference to a named game.
One game in 6,720 is 0.015% of the sample and moves nothing on this page. Two independently compiled samples of the same seasons agreeing on every other cell to the unit is a strong check on both, and it is the reason the conditional figures here are presented with confidence while the discrepancy is presented at all.
Every published figure is re-derived from the game-level sample by an automated test, separately from the run that produced it, so a number that survives a regeneration while meaning something different still fails. The window arithmetic, the pricing arithmetic and the cross-check above are all tested.
What this does not measure
It does not observe prices that were available. The study sees one closing consensus per game, not the two books a middler would actually have needed. So it answers how often the window landed, never how often the window could have been bought, and the second question decides whether any of this is actionable. Treat every rate here as a numerator.
It assumes a price rather than observing one. The 4.76% break-even is −110 on both legs. Real middles are frequently priced worse than that on at least one side, and a worse price raises the bar quickly.
It is one league and one era. The NFL from 2000 to 2025. Scoring rates and the shape of the margin distribution have both moved within that window, and nothing here is tested for stability across it. It says nothing about college football, basketball or any other sport, where the key-number structure is different or absent.
Closing lines only. No opening lines, no intraday movement, no live prices. A middle bought early against a line that then moved is a different bet from anything measured here.
It is a count, not a forecast. These are historical frequencies over a fixed sample. They are not probabilities for a future game, and a reader treating a 26-season base rate as a prediction for Sunday is making a claim this study does not support.
Sources, data and license
What was computed, the data it came from, and what you may do with it. Everything below is here so this study can be checked rather than taken on trust.
How the figures were produced
A count, not a model. Every played NFL regular-season game from 2000 to 2025 carrying a closing spread and a closing total, 6,719 of them, with each game's final margin and total compared against the middle window sitting symmetrically on its own closing number. There is no simulation, no seed and no randomness: the input is a fixed list of games and every figure is a count over it, so the same results come out byte for byte on any machine. The sample, the window model, the run and a suite that re-derives every published figure from the game list are the study's working files.
The data
One reported cell: a market, a closing number, a middle width, and how often that middle landed across 2000 to 2025. The file carries all 102 rows, including those this write-up had no room to show.
Download the full result setCSV, 102 rows
What is in each column
- market
- spread or total.
- closing_line
- The closing number. Spreads are home-relative: 3 means the home team was laying three points, −3 means the road team was.
- middle_width_points
- Gap in points between the two numbers the middle is bought at, which is also the count of results that win both legs.
- bet_number_low
- The lower of the two half-point numbers the middle is bought at.
- bet_number_high
- The higher of the two. Both are half-points, so neither leg can push.
- winning_results
- The results that collect both legs, semicolon separated. Margins are home-relative; totals are combined points.
- games
- Games in the sample that closed on this number.
- middles_landed
- Of those, how many finished inside the window.
- conditional_rate
- middles_landed divided by games. How often the middle landed, given the closing number that built it.
- unconditional_rate
- How often those same results occurred across the whole sample, ignoring the closing number.
- unconditional_two_way_rate
- Spreads only: the same window counted without regard to which side won by it, the ceiling an unsigned published margin table gives. Reported as n/a for totals, where there is no side to name.
Where the data and the comparisons came from
One public data file behind every figure on this page, the reference manual that defines the two price fields it is read through, and three published sources this study checked itself against rather than drew from. actfap has no relationship with any operator, sportsbook or data provider, and a citation here is provenance, not endorsement.
The commercial betting-content sites among them carry no link, and that is deliberate. We link to the page a claim was read from when that page is a source of record. A site running offer calls to action is not, so it is named in full instead, with what we read from it. A citation you cannot click is weaker than one you can. It is stronger than sending a reader through a signup funnel to check our arithmetic.
nflverse · nfldata, data/games.csv
Closing spread, closing total and final score for every NFL game, the file this study's sample is built from.
https://raw.githubusercontent.com/nflverse/nfldata/master/data/games.csv
Read directly from the publisher · retrieved 2026-09-11
The exact snapshot used is pinned by SHA-256 in the study's working files. The file is updated during the season, so a later retrieval will not match it.
The Comprehensive R Archive Network · Package 'nflfastR' reference manual
The definitions of
spread_lineandtotal_lineas closing prices sourced from Pro-Football-Reference, which is what licenses this study to call them closing lines.https://cran.r-project.org/web/packages/nflfastR/refman/nflfastR.html
Read directly from the publisher · retrieved 2026-09-11
OddsIndex · Middle Betting Strategy: How to Find and Bet Middles
The approximate hit-rate ranges by middle width that circulate in betting guides, quoted in Finding 3 as the reading this study measures against.
Read directly from the publisher · retrieved 2026-09-11
Not linked: A commercial betting-content site. We cite what it publishes rather than send a reader to it, on the same rule that governs every outbound link on this site.
BettingUSA · Middle Bets Explained: How Middling Works and When To Try It
Publishes the same break-even arithmetic this study uses, 4.76% on two legs at −110, arrived at independently. Read as a check on our own figure.
Read directly from the publisher · retrieved 2026-09-11
Not linked: A commercial betting-content site carrying offer calls to action. Cited for its arithmetic, which a reader can re-derive from this page without visiting it.
Clean Up Hitter · NFL Common Scores
The independently compiled margin and total frequencies for the same 26 regular seasons, used as this study's cross-check.
https://cleanuphitter.com/nfl/stats/nfl_common_scores.php
Read directly from the publisher · retrieved 2026-09-10
Read on the retrieval date shown and transcribed then for the middle calculator's unconditional table. This study compares against that transcription and did not re-read the page.
Last reviewed
September 11, 2026
Next expected update
The 2026 regular season completes in early January 2027 and adds a 27th season to the upstream file. Nothing about this study expires before then: it is a count over closed seasons, not a live figure.
Follows nflverse, which updates games.csv through each NFL season.
Reuse and license
You may copy, republish, adapt and build on this study's data and findings, including commercially, as long as you credit actfap and link back to this page.
Released under the Creative Commons Attribution 4.0 International license. If you find an error in the data, we would rather hear about it than not.
How to cite this
Korsager, A. (2026). The number the market hung. actfap. https://actfap.com/research/nfl-middles-2026
Questions this study answers
- How often does an NFL middle actually hit?
- Across 6,719 NFL regular-season games from 2000 to 2025, a one-point spread middle built on a whole-number closing line landed 4.80% of the time and a one-point total middle landed 2.88% of the time. Widening to two points on a half-point closing line raises those to 8.22% and 5.68%, but needs the two books two full points apart. Break-even on two legs at −110 is 4.76%.
- Why is the usual key-number table wrong for a spread middle?
- Because it counts a margin in both directions and a spread middle names a side. With the market closing at Home −3, the home team won by exactly three in 10.59% of games, against the 14.99% the two-way table reports. When the road team is the favorite the gap is wider still: 7.54% against the same 14.99%, an overstatement of 2.0 times.
- Does the closing line tell you where a game will land?
- Much less than people assume, and much less for totals than for spreads. Knowing the closing spread removes 18.6% of the variance in the final margin; knowing the closing total removes 10.3%. For the tightest total middle, conditioning on the market's own number moved the hit rate by 0.06 of a percentage point over 26 seasons.
- Which NFL spread is the best one to middle?
- Of the whole numbers the market hangs often, three was by a distance the most productive: a one-point middle on Home −3 landed 10.59% of the time. Four and one were close to dead, at 0.99% and 0.73%, and two reported whole numbers landed zero one-point middles in 26 seasons, which at their own sample sizes is not statistically remarkable. This is a historical frequency rather than advice: whether any of them clears its price depends on the two prices on offer, which this study does not observe.
Put it to work
Run your own numbers through it
- Middle Calculator Price a middle on both sides of a moved line: what it pays, what a miss costs, and how often it has to land.
- Arbitrage Calculator Check whether two or more prices on the same event lock in a profit, and get the stake for each side.
- CLV Calculator Work out your closing line value, and whether the sample you have is large enough to mean anything yet.
More original research
See every study- 0 of 10price buckets where the mispricing was bigger than what it cost to reach it
Oct 2026 · 11 min read
The longshot gap, and the spread that swallows it The academic literature reports that prediction-market longshots are mispriced. We checked it on 12,755 settled Kalshi contracts, priced a day before they closed, and then did the part nobody had done: measured what reaching the mispricing would have cost at the same price, at the same moment, out of the same rows. In 10 price buckets out of 10, the cost was larger than the gap. - +1.35 pointsof the +1.38 points rise in Nevada's slot hold is the same machines getting more expensive, not players moving to worse ones
Sep 2026 · 12 min read
Hold is not the house edge Nevada publishes what its casinos actually keep, every month, and has since 2004. We read all 269 reports as one dataset. Two of the numbers in them look alike and measure different things, and almost everything people get wrong about this data starts there. - 9.9cmedian two-way cost per dollar of payout across 78,119 quotable Kalshi markets
Sep 2026 · 9 min read
What a prediction market actually charges Prediction markets are sold on not charging vig. We snapshotted every quotable contract on Kalshi and priced the two-way cost the way a sportsbook's hold is priced. On 78,119 markets the median came to 9.9c per dollar of payout, against 4.8c at a standard sportsbook line. On the markets people actually trade, the answer flips.
21+. Available only where regulated.
This is research into how often a middle landed, not advice to bet one. actfap carries no affiliate links, no ads, no paid placements and no operator relationships of any kind.
Gambling can cause serious harm. If it is affecting you or someone you know, the National Problem Gambling Helpline is free, confidential and open 24/7, call or text 1-800-MY-RESET (1-800-697-3738), or chat at ncpgambling.org/chat. More on staying in control.