Sportsbooks post a "tackles + assists" line for defenders — the player's combined solo
and assisted tackles in the official box score. It is a volume stat driven by snaps, role,
how many plays the opposing offense runs and where the ball goes, which makes it more
predictable than most props and more sensitive to game script than books tend to price.
The three-stage model
Stage one is a LightGBM regression trained on 2021-present player-game data that
projects each defender's expected tackle count from role (snap share, tackles per snap,
every-down status, depth-chart rank, and PFF alignment — how often the player lines up in
the box, the slot, at corner, deep or on the line, plus stop rate, missed-tackle rate and
tackling grade), the opponent's offense (pace and neutral-script pace, pass rate and
early-down pass rate, rush direction, depth of target, middle-of-field and tight-end
target share, quarterback time to throw, and personnel — how many receivers, tight ends
and backs it puts on the field), the defense's own context (plays faced, personnel shares,
vacated snaps) and the game environment (spread, total, weather, rest). Stage two wraps that mean in a
negative-binomial distribution fitted per position group, so the spread of outcomes is
realistic rather than assumed. Stage three passes the resulting P(over) through an
isotonic calibration curve fitted on held-out weeks, so a stated 60% means 60%.
The two pick gates
A projection becomes a pick only when it clears both gates at once: the projection
must sit at least 0.75 tackles from
the line, and the calibrated probability for that side must beat the book's vig-free
implied probability by at least 4.0 points. One without the other is a lean, not a pick. Picks are priced at
-160 or longer.
Filters on this page
"Min edge" hides picks whose projection-vs-line gap is smaller than the number you type
(it starts at the model's own gate). The odds range keeps only picks priced inside your
band — tighten the low end if you never take heavy juice, raise the high end to see
long-shot unders. The two filters are independent and the counter shows how many of the
week's picks survive.
Lock and inactives
Picks are projected during the week as lines move, then locked
90 minutes before kickoff, when official inactives post.
The locked line, side and price are what gets graded; anything you see before that is a
projection that can still change. Players ruled out before lock are dropped, not graded.
Void rules
A pick is void — no win, no loss, no units — if the player logs zero defensive snaps,
is inactive, or the game is postponed. One or more snaps means the bet has action and is
graded on the official box score, whatever happened after. Pushes (actual exactly equal
to the line) are recorded separately and don't count toward the record.
Reliability tiers and "low sample"
Every defender gets a reliability tier from how well the model has tracked them
specifically — accuracy against actual tackles, whether it runs persistently high or low,
and whether its probabilities have been honest. Elite and Reliable players have a clean
track record; Watch players have a wobble; Unreliable players are suppressed from picks
entirely. "Low sample" flags a player with fewer than 8 graded games
for the model — the pick still cleared both gates, but there is not yet enough history to
say how well the model knows this player.
Rookies and first-time starters
A defender with no prior NFL game never becomes a pick, whatever the projection says. With no snap or tackle history the model
leans on weaker signals — years of experience, draft position and the team's published depth
chart — and a book posting a line for such a player usually means it expects a starter. The
projection is still shown; the pick is withheld until the player has real snaps on record.
How results are counted
The record, units and ROI on this page count only line-backed picks that were
published here and then graded — one flat unit per pick, at the locked price. Backtests,
projections without a book line and voids are never mixed in. "Avg CLV" is closing line
value: how far the market moved toward our side between the lock and kickoff, in
probability points. Beating the close consistently is the best early evidence the model
is real; a few weeks of win-loss is mostly noise.
Player profiles and pick breakdowns
Click a player's name (on a card or in the results table) for his profile:
his last 10 games against the closing line each week, hit rates at today's line over his last
5, last 10 and this season, his role (snap share, depth chart, where PFF has him lining up), the
matchup (how many tackles this opponent gives up to his position group, pace, pass rate,
spread and total) and how accurate the model has been on him. Games he missed or played zero
defensive snaps in are left out — a DNP is not a zero.
Click anywhere else on a card (or a results row) for "Why this pick": the
projection against the line, the model's probability against the book's vig-free
probability, the full spread of outcomes the model expects, and the five factors that moved
this projection most, in plain English. Each factor's "≈ tackles" effect is an approximation
— the model's factors multiply together rather than add, so they won't sum exactly to the
projection. Very large probability edges carry a warning: in testing they won less often than
moderate ones.
Record48-23-0
Voids0
Units+13.80
ROI+19.4%
Avg CLV-0.2 pp
Graded picks71
Picks — Week 7
No picks match these filters.
Results
Week
Picks
W-L-P
Units
ROI
2026 w6
13
10-3-0
+4.40
+33.8%
2026 w5
11
5-6-0
-2.04
-18.6%
2026 w4
10
5-5-0
-1.50
-15.0%
2026 w3
9
5-4-0
+0.13
+1.5%
2026 w2
22
18-4-0
+9.75
+44.3%
2026 w1
6
5-1-0
+3.06
+51.0%
Wk
Player
Pick
Line
Proj
Actual
Result
Units
CLV
2026 w6
WAS vs SF
UNDER -155
5.5
4.4
3.0
Won
+0.65
—
2026 w6
LV vs BUF
UNDER -110
7.5
6.5
6.0
Won
+0.91
0.0 pp
2026 w6
LV vs BUF
UNDER -145
3.5
2.5
1.0
Won
+0.69
0.0 pp
2026 w6
KC vs LAC
UNDER -145
5.5
4.5
2.0
Won
+0.69
—
2026 w6
IND vs TEN
OVER -155
3.5
5.1
6.0
Won
+0.65
—
2026 w6
BAL vs CLE
UNDER -145
4.5
3.7
2.0
Won
+0.69
0.0 pp
2026 w6
ATL vs CHI
UNDER -140
5.5
4.2
7.0
Lost
-1.00
0.0 pp
2026 w6
IND vs TEN
UNDER -130
4.5
3.5
2.0
Won
+0.77
0.0 pp
2026 w6
CLE vs BAL
UNDER -155
3.5
2.7
3.0
Won
+0.65
+0.3 pp
2026 w6
CLE vs BAL
UNDER -140
5.5
4.3
1.0
Won
+0.71
0.0 pp
2026 w6
TB vs PIT
UNDER -150
5.5
4.3
6.0
Lost
-1.00
—
2026 w6
HOU vs JAX
UNDER -145
5.5
4.3
8.0
Lost
-1.00
+0.3 pp
2026 w6
JAX vs HOU
OVER +100
4.5
5.5
6.0
Won
+1.00
0.0 pp
2026 w5
BUF vs LA
UNDER -115
4.5
3.7
10.0
Lost
-1.00
0.0 pp
2026 w5
SF vs SEA
OVER -155
3.5
4.9
2.0
Lost
-1.00
0.0 pp
2026 w5
ARI vs DET
OVER -125
6.5
7.7
3.0
Lost
-1.00
0.0 pp
2026 w5
NYJ vs CLE
UNDER -150
3.5
2.6
3.0
Won
+0.67
0.0 pp
2026 w5
LV vs NE
UNDER -120
8.5
6.8
4.0
Won
+0.83
0.0 pp
2026 w5
NYG vs WAS
UNDER -150
6.5
5.7
4.0
Won
+0.67
—
2026 w5
CLE vs NYJ
UNDER -145
5.5
4.4
5.0
Won
+0.69
0.0 pp
2026 w5
TEN vs HOU
OVER +110
4.5
5.5
6.0
Won
+1.10
-1.3 pp
2026 w5
WAS vs NYG
UNDER -150
5.5
4.7
7.0
Lost
-1.00
0.0 pp
2026 w5
JAX vs PHI
UNDER -120
7.5
6.4
10.0
Lost
-1.00
-1.0 pp
2026 w5
PHI vs JAX
UNDER -145
8.5
7.5
11.0
Lost
-1.00
-3.4 pp
2026 w4
ATL vs NO
OVER -150
4.5
6.2
8.0
Won
+0.67
—
Last 25 graded picks. Results count line-backed picks only — one flat unit each at the locked price. Picks tagged as-of were projected before kickoff by the week backfill and graded against the closing line rather than locked live.
