🏈 NFL Betting Trends — Week 4

This page is a shortlist of things worth a second look, not a list of picks. It tests hundreds of situations every week, and when you test hundreds of anything, some of them look remarkable for no reason at all. Everything below exists to help you tell those apart.

How a row is put together

The ticket

The bet the record is about — the team and the number you would have had to lay or take.

ATS

Which market. ATS is against the spread; Over and Under are the total. A spread bet is graded for each side separately, because backing the favourite and backing the underdog are two different bets. A total has no side, so it is graded once per game.

The sentence

The situation being described — the filter that picked these games out of the 6,281 in this page, 2003 to 2026.

win rate

How often that bet won, ignoring pushes. 52.38% is break‑even, not 50%, because a losing bet at −110 costs more than a winning one pays. Anything under that number lost money.

record

Wins–losses–pushes. This is the number that decides how much the win rate is worth, and it is the one most easily skipped past.

roi

What a flat stake on every one of those games would have returned, at −110. Break‑even is 0%.

avg cover

By how much, on average, per game — not just how often. A trend that wins narrowly every time and one that wins in blowouts have the same win rate and are not the same thing.

🔥

How strong the evidence is. This is the important one, and it is explained next.

Why the fire is not the win rate

The glyph is not how often it won. It is how far the record sits above break‑even relative to how many games it is built on — what statisticians call sigma, and what you can read as “how hard would this be to get by luck”.

Here are the same two trends. The first has the better win rate. The second is the better read, and it is not close:

68.0% off 25 bets is 1.6σ. 61.1% off 190 bets is 2.4σ. Seven points of win rate, and the longer record still wins.

• 🔥 — 2σ or more. Strong. Still not a guarantee, and still one of many things tested.
• 📈 — 1 to 2σ. Moderate. Interesting, unproven.
• 🤷 — under 1σ. Weak. Shown so you can see what weak looks like next to the rest.

This is why the page sorts by sigma and never by win rate. A short record with a spectacular percentage sorts below a long one with a modest edge, on purpose.

What “cooling” means

A record going back to 2003 can be carried entirely by seasons that are over. A trend is tagged cooling, and capped below 🔥, when either of two things is true:

• it has lost money over the last 3 seasons, or
• it is under 50% across the last 10 times it actually came up, whatever the career number says.

Both need at least 8 recent decisions before they are allowed to fire, so a handful of games cannot convict a thousand. They print the same word deliberately — to a reader they mean one thing, “this is not live” — and hovering the tag says which one fired.

Open a row

Every row on the board opens. Inside is the last 20 times that situation came up, game by game, won or lost. It is the fastest way to see whether a record is a steady thing or one hot patch with a long tail, and it is worth doing before trusting any row on this page.

What it looks at

Each trend stacks up to 3 conditions drawn from these:

• the opening week, the last week, or the playoffs
• early, middle or late in the season
• home or away
• favourite or underdog, and by how much
• division games
• how the team has been playing lately
• whether the team is coming off an overtime game
• whether the team is still without a win and a cover this season
• days of rest, including off a bye
• an unusually light or heavy travel load
• time zones crossed to get there
• indoors or outdoors
• the kickoff time
• how high the total is
• temperature at kickoff
• wind speed at kickoff
• how far the weather is from what that team is used to
• whether either team is coming off an overtime game on a short week

Every combination of up to 3 of them is tried, against every game on the slate. That is where the hundreds of situations come from, and it is also the whole problem.

What gets thrown away, and why

Most of the work here is subtraction. A wide scan manufactures impressive-looking rows, so:

• Nested duplicates are merged. If “home underdogs” and “home underdogs, outdoors” say the same thing, the narrower one only survives if it genuinely beats the broader one — on win rate and on sigma. Otherwise you would read one finding four times and feel four times as sure.
• A subject may not contradict itself. Two cuts of the same team once pointed at opposite sides of the same game. When that happens now, both are removed — picking a winner between two halves of an even record is the mistake, not the tie.
• Dead ends stay out. Conditions that were measured and predicted nothing are not offered as dimensions, however good a story they make.

The part worth taking seriously

The whole scan was re-run on randomly flipped results — the same conditions, the same slate, winners decided by a coin. It finds about as many extreme-looking records as the real data does.

That is not a flaw being confessed; it is the fact this entire page is built around. The number of remarkable records you find is mostly a property of how many situations you tested, not of how much edge is out there. Consolidation, the contradiction rule, the cooling tags, the sigma sort and the 1σ floor all exist to stop that arithmetic reaching you as false confidence.

So: use a row as a question to go and answer — who is hurt, what is the number now, does the situation still apply — and never as an instruction.

Nothing here is betting advice, and no result on this page is a prediction. Records are graded at −110 regardless of the price actually available, so real returns will differ. 21+. If gambling stops being fun, stop — help is available at 1‑800‑GAMBLER.