Why NFL Week 1 and 2 player props are unreliable
Every NFL projection you read in Week 1 is mostly about last season, because this season has not produced anything to project from yet. No snap counts, no target shares, no touches. That is true of our model and of every other one. What differs is what each of us does about it.
This season so far, we produced 877 NFL player-prop reads and showed you 15. The other 862 were written down, time-stamped and graded, and never put in front of anybody. The numbers on this page come out of that record.
What actually breaks in the first two weeks
A player prop projection is built from how a player is used. How many snaps he plays. What share of the targets he sees. How deep he runs. How often his team hands him the ball near the goal line. None of that exists in Week 1. The season is zero games old.
So a model has two choices, and both are bad. Use last season, and you are projecting a player whose role may have changed completely over an offseason. Use nothing, and you have no projection at all.
Roles move more than people expect. A rookie takes the slot. A back loses third downs. A coordinator arrives and throws forty times a game. Last season is genuinely informative about the player and genuinely misleading about the job he now has, and in September you cannot tell which is which.
What we tried, and what it measured
The repair is a standard one: take each player’s rate from last season, pull it toward his position’s average, and weight that blend by how much data supports it. A player with a long, stable history moves a little. A player with almost no history moves a lot, toward the average of players like him. As this season fills in, the blend hands back to the real numbers.
We fitted that on two seasons and tested it on a third it had never seen. The result is the reason this page exists, and it is not the clean win we wanted.
| Projection | Calibration error | Verdict |
|---|---|---|
| Passing yards and passing TDs | 6.3 points | Good enough to show |
| Rushing yards | 11.2 points | Not good enough |
| Receiving yards and receptions | 14.1 points | Not good enough |
Calibration error is the average gap between the probability a model states and how often that thing then happens. Lower is better. Our own baseline for a usable read is about 6.8 points, so passing cleared it and the other two did not.
Two other things fell out of that test, and both are worth knowing if you are building your own numbers.
The blend has to know when to let go. Our first version kept leaning on last season long after this season had plenty of data of its own. By late in the year it was moving probabilities as much as 26 points away from what the current numbers said. A prior is a bridge, not a destination.
Rushing is not a fixable term, it is variance. The same method scored 7.6 points on one pair of seasons and 11.2 on another. A number that unstable across seasons is telling you the early-season rushing signal is genuinely noisy, not that one more adjustment would rescue it.
The thing that surprised us
Our receiving projection estimates yards per reception from how deep a receiver runs, rather than from his actual yards per reception. That looks like a shortcut, so we tried replacing it with the measured number. Calibration got worse at every setting we tried.
The reason is the whole lesson of this page in one example. How deep a receiver runs is a stable fact about his role. His yards per reception over ten or twenty early-season catches is mostly noise. The rough number that is stable beat the exact number that is not, which is a trade you will keep meeting in September.
What we do about it, in public
We rate the whole board, then withhold the parts our own test says are not calibrated yet. Passing reads carry a grade. Rushing, receiving and touchdown reads are produced, stored and graded, and their grades are not shown.
| Market | Shown | Withheld |
|---|---|---|
| Anytime touchdown | 0 | 432 |
| Receiving yards | 0 | 159 |
| Receptions | 0 | 156 |
| Rushing yards | 0 | 77 |
| Rushing plus receiving yards | 0 | 38 |
| Passing touchdowns | 12 | 0 |
| Passing yards | 3 | 0 |
Read from our public NFL record, 2026 season to date. Withheld reads are still written before kickoff and still graded.
Withholding a read is not free. Week 1 is the biggest week of the NFL season and holding back most of a board is a bad way to sell subscriptions. We do it because a grade is a claim about a probability, and we do not have one worth making on those markets yet.
The withheld reads are still graded, which is the part that matters later. When the season is over we can check what those suppressed grades would have done, and either earn the right to show them next September or find out we were right to hide them.
How to read early-season props yourself
- Ask what changed about the job, not the player. New coordinator, new depth chart, a rookie drafted at his position. Role changes are what break last season’s numbers.
- Weight opportunity over production. A receiver who ran a full route share tells you more than one who caught a long touchdown. Opportunity stabilises in about two or three games. Production does not.
- Distrust any confident number in Week 1, including ours. If a tool gives you a crisp probability on a Week 1 rushing prop, ask what it computed that from. There is no this-season usage for it to have used.
- Passing is the exception, and it has a reason. Quarterback volume is set mostly by scheme and pace, which carry over between seasons far better than a running back’s workload does.
- Wait three weeks if you can. That is roughly where our own reads stop leaning mostly on the prior year, and it is free to wait.
Questions
Are NFL Week 1 player props worth betting?
They are the hardest props of the year to model, and that cuts both ways. Every projection in Week 1 leans on last season, because this season has produced no snap counts, no target shares and no touches yet. Our own replay found passing projections survive that and rushing and receiving projections do not, so we show passing reads in the opening weeks and withhold the rest.
Why do NFL projections get better after Week 3?
Because the model finally has this season to look at. Roles change over an offseason in ways last season cannot tell you: a rookie takes the slot, a back loses third downs, a coordinator throws more. Two or three games of real usage is roughly where our reads stop leaning mostly on the prior year.
Can you use last season to predict Week 1?
Partly, and it is worth doing. We blend each player’s prior-season rate toward his position’s average, weighting by how much data he has. Tested on a held-out season, that got passing yards to a 6.3-point calibration error against a 6.8-point baseline. The same method left rushing at 11.2 and receiving at 14.1, which is not good enough to publish a probability on.
Does a big Week 1 game mean anything?
Very little on its own. One game is one game, and early-season lines move hard on exactly those performances. The question worth asking is whether the role changed, not whether the yards were high. A receiver who ran a full route share in Week 1 has told you more than a receiver who caught one long touchdown.
What does ET Parlays do about it?
We rate the whole board and then withhold most of it. Every read is still written down before kickoff and still graded, so we can check ourselves against the reads we did not show. The counts on this page are read from that record.
Nothing here is betting advice. Bet only what you can afford to lose. If it stops being fun, take a break.