Favorites did not quietly print covers over the last 27 completed regular seasons. They went 3,298-3,450-189 ATS from 1999 through 2025, a 48.9% cover rate over 6,748 decisions. The season line bounced from 58.9% at the top to 43.4% at the bottom. That is a wide range wrapped around a long-run record with more losses than wins.
This is the record-first answer to “how often does the favorite cover?” It is not “about half” and it is not one cherry-picked season. It is a season-by-season ledger graded at the closing spread. For the mechanics of favorite, underdog, cover, and push, start with how NFL spreads work. Then keep the chart open while you read the annual swings.
The full 1999–2025 result
The snapshot contains 6,967 played regular-season games in the selected seasons. Thirty closed at pick’em and therefore had no favorite. The remaining 6,937 games produced the favorite record above. There were 189 pushes, so the cover-rate denominator is 3,298 wins plus 3,450 losses, not all 6,937 games.
Nineteen of the 27 seasons finished below 50% for favorites. Eight finished above 50%. The above-50 seasons were 2003, 2005, 2007, 2013, 2016, 2017, 2023, and 2024. That list is descriptive. It is not a sequence a bettor could have known before each season began.
Every season in W-L-P form
| Season | Favorite ATS | Cover rate |
|---|---|---|
| 1999 | 110-125-9 | 46.8% |
| 2000 | 112-124-7 | 47.5% |
| 2001 | 110-124-11 | 47.0% |
| 2002 | 110-139-5 | 44.2% |
| 2003 | 123-121-9 | 50.4% |
| 2004 | 122-124-6 | 49.6% |
| 2005 | 142-99-10 | 58.9% |
| 2006 | 108-141-7 | 43.4% |
| 2007 | 130-121-5 | 51.8% |
| 2008 | 123-127-6 | 49.2% |
| 2009 | 121-127-8 | 48.8% |
| 2010 | 123-128-5 | 49.0% |
| 2011 | 121-124-11 | 49.4% |
| 2012 | 119-132-5 | 47.4% |
| 2013 | 135-115-6 | 54.0% |
| 2014 | 120-130-6 | 48.0% |
| 2015 | 114-134-8 | 46.0% |
| 2016 | 128-122-5 | 51.2% |
| 2017 | 132-113-8 | 53.9% |
| 2018 | 114-133-9 | 46.2% |
| 2019 | 113-133-10 | 45.9% |
| 2020 | 114-142-0 | 44.5% |
| 2021 | 125-143-4 | 46.6% |
| 2022 | 118-143-10 | 45.2% |
| 2023 | 137-121-14 | 53.1% |
| 2024 | 143-125-4 | 53.4% |
| 2025 | 131-140-1 | 48.3% |
The 2005–2006 warning label
The most useful pair is not 2023 and 2024. It is 2005 and 2006. Favorites covered 58.9% in 2005, the highest season in the file. One year later, they covered 43.4%, the lowest. The side did not merely drift back toward the middle. It crossed the entire observed range in consecutive seasons.
A bettor building a “favorites are hot” rule after 2005 would have selected the weakest favorite season in the snapshot next. A bettor building an automatic fade after 2006 would have watched favorites recover to 51.8% in 2007. The annual record is valuable because it shows that the label “favorite” is not a stable edge by itself.
The same caution applies to the newest run. Favorites posted 53.1% in 2023 and 53.4% in 2024. In 2025 they fell to 48.3%. The point is not that 2025 disproved the previous two years. The point is that a side-level rate can move several percentage points without creating a reusable model.
What this record does and does not say
It says how the favorite performed at the final spread stored in the snapshot. It does not say when the favorite became the favorite. It does not preserve an opening spread. It does not say whether a bettor laid a different number on Monday and watched it close elsewhere. Those questions belong to line-history data, not to a closing-line result table.
It also combines every favorite size. A small favorite and a large favorite both add one decision. That is correct for the broad question in the title. It is not enough for a price-sensitive strategy. A proper follow-up would predeclare spread buckets, keep each bucket’s record and sample, and test it outside the period used to choose the buckets.
Finally, the record does not claim profit. The source includes spread prices in many games, but the article does not impose a single price or staking rule on a 27-season count. ATS W-L-P is the clean common denominator. Profit requires the actual price attached to each side and a declared unit rule.
The Receipts Drawer
The long record is plain: favorites covered 48.9% of decisions. The best season was 2005. The worst was 2006. Only eight seasons finished above 50%. Those receipts are useful because they make a popular shortcut harder to say with a straight face: “just take the better team.” The closing spread already charges for the better team.
Use the Analytics surface to inspect records with the market and time window named. Use the NFL picks page for current graded picks. If a current angle is described as “favorites have been covering,” ask for the exact start date, W-L-P line, push rule, and closing-line source before the percentage gets a vote.
How to read the chart
The y-axis is the favorite’s cover rate, not its straight-up win rate. A favorite can win the football game and fail to cover. Each point uses only non-push decisions for the percentage. The record table preserves pushes because a refund is still part of the audit trail.
The line is not a smooth estimate of a hidden law. It is 27 separate regular-season samples. Season sizes differ across the table, so later rows can carry more games than early rows. That makes W-L-P more informative than a percentage printed alone. The chart shows volatility. The table shows what produced it.
What a useful baseline looks like
A baseline should be broad enough that a current split cannot choose it after the fact. The 1999–2025 favorite record uses every completed regular season in the snapshot, one closing convention, and one grading formula. It is not tuned to a team, spread range, month, or television window.
That makes it a reference point rather than a strategy. A current model that selects favorites should compare its own recorded favorite picks with the broad favorite baseline, but only on games the model actually selected. It should not claim the league-wide 48.9% rate as the expected return for its narrower sample. Selection changes the population.
The baseline also keeps a reviewer honest about direction. If a post says favorites were strong, the table can show whether the claim refers to 2023, 2024, both, or the full archive. If the date range disappears from the sentence, the sentence has lost part of its fact.
A future update should append the next completed regular season to the same ledger. It should not replace 1999 with a later start merely because the newest window looks cleaner. A separate recent split is fine when its dates stay in the label. The full baseline remains the audit trail.
That practice also makes revisions visible. If a source corrects a final score or closing line, the snapshot hash changes and the affected season can be recomputed. The article should then update its date and explain the changed input instead of silently editing the percentage.
Where these numbers come from
How we counted: We concatenated the four files matching data/training-snapshots/nfl/games-*.csv. We kept played rows with game_type = REG, seasons 1999–2025, and a closing spread_line. The dictionary defines positive spread as the home team favored. Favorite ATS margin is result - spread_line when the home team is favored and its negative when the away team is favored. Pick’em rows with spread_line = 0 were excluded because they have no favorite. Positive margin is a cover, zero a push, negative a loss. Cover rate is wins divided by wins plus losses. Snapshot identity and shard hashes are in data/training-snapshots/nfl/manifest.json.
NFL ATS cover-margin distribution
Distribution of (final margin − closing spread) across an NFL season. Roughly normal with mean ≈ 0 and standard deviation ≈ 13 points, which is why most ATS edges live in the ±1.5 point window.
Model calibration: predicted vs observed
Predicted win probability bucket vs the empirical win rate inside that bucket on the test set. Points on the y=x reference line are perfectly calibrated; points below mean the model is overconfident in that bucket.
