High Totals Land Under. Low Totals Don't Land Over.

There's a tidy theory that extreme totals regress: when the number is sky-high, expect the game to fall short; when it's rock-bottom, expect points to sneak over. It's symmetric, it's intuitive, and it's exactly the kind of thing that sounds true.

It's half true. We checked every NFL game since 2003, and the data splits the theory cleanly down the middle: the high side is real and stable; the low side is a mirage that reverses when you test it honestly. Getting there also required fixing a subtler mistake that trips up almost every "high total" analysis you'll read.

First, "high" has to mean high for its era

A total of 49 was a shootout number in 2004 and an ordinary Sunday number in the 2020s — league scoring has drifted up over two decades. So classifying games against a fixed line ("49 or higher = high total") quietly lumps 2004 shootouts together with 2022 average games, and scoring inflation masquerades as a pattern.

The fix is to measure each game's total against that season's average. A game 7+ points above its season's norm is genuinely high for when it was played; one 7+ points below is genuinely low. (This is why the League Trends totals view now uses season-relative bands instead of the old fixed 49/39 thresholds — the fixed version was measuring inflation as much as anything.)

The all-era gradient looks like clean regression

Bucket every game by how far its total sat from the season average, and the under rate climbs in a straight, satisfying line:

Low totals lean over (47.1% under = 52.9% over); high totals lean under (54.7%). Monotonic, symmetric, exactly what the regression theory predicts. If you stopped here — and most analyses do — you'd conclude both tails regress.

Splitting by era detonates the low side

The honest test is whether a pattern holds on data it wasn't discovered on. Split the sample into an earlier stretch (2003–2021) and a recent hold-out (2022–2025):

The high-total→under leg is stable — it leans under in both eras and actually strengthens (53.7% → 60.0%). The low-total→over leg completely reverses: the earlier era's 55.9%-over pattern flipped to 58.2% under in the hold-out. The "low totals go over" half of the theory wasn't a pattern at all — it was a coincidence of one era that vanished the moment it was tested forward.

So the real, defensible finding is asymmetric: unusually high totals lean under and have kept doing so; unusually low totals do not reliably do anything.

Why the asymmetry makes sense

High totals are set for games the whole world expects to be shootouts — big offenses, dome roofs, soft defenses, revenge-game storylines. Expectations of fireworks run hottest for exactly these matchups, and the games, on average, don't quite deliver: 60% of the most-inflated totals finished under in the recent window. The market isn't broken, it's optimistic, and the over pays the optimism tax.

Low totals are the opposite psychology — nobody hypes a defensive slugfest, so the number is set soberly, with no enthusiasm premium to correct. There's simply less error to exploit, which is why the low tail behaves like the efficient middle.

The honest caveats

• The tails are thin. ~390 games in the high band over two decades, fewer in the recent hold-out. The direction is consistent, but small samples mean the exact percentage carries real uncertainty.
• This is a descriptive pattern, not a system. It's a lens for reading a slate, not a rule that prints results — 60% under still means the over hits 2 games in 5. We tested whether "high total" could improve our prediction model as an input and it couldn't; like our weather findings, it's a conditional read, not a factor a model can use across the whole schedule.
• Season-relative is doing real work. Every number here depends on comparing a total to its own era. Any version of this analysis built on fixed thresholds is partly measuring the last 20 years of scoring inflation.

See it by season

The season-relative totals bands are on the SpreadTrends League Trends page under Totals Trends — every NFL game since 2002, bucketed against its own season's scoring level so a 49 is judged by the year it was played in.

Data: 2003–2025 seasons; totals measured as deviation from the season average; earlier era 2003–2021, hold-out 2022–2025. Games landing exactly on the total are excluded from percentages.