NFL
More
Record · 2026
Our picks
29–17
63.0% of 46 games
Market pick
31–15
same games · 67.4%
Graded
Weeks 1–3
weeks graded so far
Confident picks
67%
hit rate on 75%+ picks (9)
5
longest streak · current 0
11.1
avg. margin error (pts)
11.4
avg. total error (pts)
2–4
when we differed from the market · the market 4–2
65.0%
backtest 2010–2025 · the market 66.5% →
Record by kickoff slot and by pick
The same record, cut the ways models usually show a bias. the market = the market pick at kickoff over the same games. Small samples early; judge after a month.
| Kickoff | Ours | % | the market |
|---|---|---|---|
| Thursday night | 1–1 | 50% | 1–1 |
| Sunday early (1:00) | 15–10 | 60% | 16–9 |
| Sunday late (4:00) | 9–4 | 69% | 10–3 |
| Sunday night | 3–0 | 100% | 2–1 |
| Monday night | 1–2 | 33% | 2–1 |
| Our pick was… | Ours | % | the market |
|---|---|---|---|
| picked the home team | 21–11 | 66% | 21–11 |
| picked the away team | 8–6 | 57% | 10–4 |
| picked the market pick | 27–13 | 68% | 27–13 |
| picked the other side | 2–4 | 33% | 4–2 |
If one unit had been placed on every pick
A thought experiment, not a recommendation and not something we do. One flat unit on each of our picks at the final market price, next to one unit on the market pick at kickoff every time. Positive means the season would be ahead; the backtest says to expect both curves below zero over time, because the market's final price already prices what we know. Shown so nobody has to wonder.
-1.8
units, our picks · 29–17 · ROI -3.8%
+0.6
units, market pick at kickoff · 31–15 · ROI +1.3%
46
games with a final market price
Payouts use the nflverse final market price: a winner at −150 returns 0.67 units, at +130 returns 1.3 units, a loser costs 1 unit. Predictions are published to show the model works. Nothing here is betting advice.
Against the market margin and the total
The same honesty for the two other things the model produces. Our projected margin is compared with the final market margin and our projected total with the closing total, the side is fixed before kickoff, and pushes are excluded. Expect these to sit near 50%: the backtest (2010–2025) covered 49.9% of market margin and hit 50.8% of totals, which is what "no edge against the market's final price" looks like. Shown because hiding it would be the wrong lesson from the win–loss record.
19–24
against the market margin · 44.2%
20–25
over / under · 44.4%
52.4%
break-even rate at a standard −110 price
| Week | ATS | O/U |
|---|---|---|
| Week 1 | 6–8 | 6–8 |
| Week 2 | 7–8 | 6–9 |
| Week 3 | 6–8 | 8–8 |
Model vs. model + analyst layer
Since v0.4 the pick of record is the analyst-adjusted probability. The untouched model probability is stored and graded beside it, so the layer has to earn its keep. What it read and changed →
28–18
stats model alone
29–17
model + analyst (pick of record)
35
games the analyst moved
1–0
record when it flipped the pick
A 17-game season is a small sample: over the 272 games in a season the 95% margin on an accuracy figure is about ±6 points, so a season proves less than it looks. The backtest over 4,146 games is the fairer test of the method; this page is the live, pre-kickoff record.
Accuracy by week
Straight-up hit rate. Lighter bar = a week still being played. Hover a bar for the numbers.
Week 1Week 3
Does confidence mean anything?
Stated confidence vs. how often the pick was right. When we say 70%, are we right 70% of the time?
Hit rate by confidence × week
Darker = the model was right more often. A healthy board is dark at the top, lighter at the bottom, in every column.
W1
W2
W3
80%+
70–80%
60–70%
50–60%