The model

How Overline makes a pick

Every number on the board is a predicted score, then a probability, then one bet. We never invent a DraftKings price, and we never swap the card after it posts.

01
Season stats
Points scored and allowed, record, recent form.
02
Shrink
Small samples get pulled hard toward league average.
03
Ratings
Attack, defense, home field, ESPN FPI/BPI.
04
Blend DK
The posted spread and total become a prior.
05
One card
The stronger playable spread or total, plus one player prop.

The predicted score

Football and basketball are point sports. Baseball, hockey, and soccer are count sports. The ingredients are the same; the engine that turns them into a score is not.

Attack / defense

How many points (or goals) a team scores and allows versus the league average, shrunk by how many games they have played.

Shrink

Weight on a team’s own stats = games / (games + shrink constant). NFL shrinks after 6 games, so a 2–0 start only gets 25% weight. The rest is league average. A hot week-one does not look like a dynasty.

Win bump

Winning percentage nudges attack up and defense up a little, so 2–0 on a flukey score still counts as a good team — just not a 27-point team yet.

Home field

A sport-specific constant. NFL is +2.4 points, mostly given to the home attack, a little taken off the visitor.

Power index

ESPN FPI (football) or BPI (basketball) is a preseason-stable point-margin rating. Early in a season it is 55–80% of the model score. It fades as real games pile up.

DraftKings prior

If DK posts a spread and a total, we convert that into an implied score: home = (total − home spread) / 2. That implied score is blended in at 52–74%, and more when our margin and the number disagree by a lot. A 40-point mismatch does not get to overrule the book.

A worked NFL example

Bills at Chiefs, week two. League typical total 44.8 (22.4 per side), home field +2.4, shrink after 6 games, noise σ 13.6.

Season pace

KC
2–0 · 27 scored, 16.5 allowed
BUF
1–1 · 24 scored, 23 allowed

After shrink (25%)

KC
+1.5 attack · +1.2 defense
BUF
+0.4 attack · −0.2 defense

Model score

KC
22.4 + 1.5 − (−0.2) + 1.3 = 25.4
BUF
22.4 + 0.4 − 1.2 − 1.1 = 20.5

Then the market

DraftKings posts KC −3.5, total 46. That implies KC 24.8, BUF 21.3. Week two is still mostly market — about 70% DK, 30% ratings — so the board lands near BUF 21.0–24.9 KC. Predicted margin +3.9 versus a −3.5 spread. With σ 13.6, Kansas City covers about 51% of the time; the over on 46 is a coin flip. The computer posts the stronger playable number, here the spread, as long as the juice is -120 or longer.

From a score to a probability

Football & basketball

A bell curve around the predicted margin. NFL noise is 13.6 points, NBA 12, NCAAF 16.2. Cover chance is how much of that curve sits on the right side of the DraftKings spread. Over chance uses a slightly wider curve on the total.

Baseball, hockey, soccer

Independent Poisson counts. The chance of the over is the chance the two scorelines add up past the total, not a bell curve. A 1.5-goal total with 1.8 expected goals is a lean, not a lock. Soccer keeps a 0–6 score matrix on the game page.

UFC, tennis, golf, F1

No score. Win probability from records, form, and (when posted) the moneyline. Golf cups use match-play scoring, not stroke totals.

Live games

Take the current score, add the leftover pregame scoring, shrink the noise with the clock. A 17–3 NFL game in the fourth is not re-handicapped from scratch.

The game page runs about 1,600–1,800 seeded simulations so the histograms and 80% intervals stay still if you refresh. The desk brief is optional and user-triggered — it never runs on page load.

Two books, not one

The probability bar is a winner lean. It is not the graded pick. Graded cards live on two ledgers:

Computer · spread or total

One numbers play

Compare the model’s cover chance to its over/under chance.
Keep only DraftKings prices at -120 or longer. No −150 favorites.
Post the stronger one. Ties prefer spread, then total, never the moneyline.
Lock that player-less card on first post. Later desks cannot move the line or the side.

Grok Expert · player prop

One player, one market

The posted DraftKings line is the median. A generic role (every QB, every assist) barely moves it.
A real player rate — season yards, points, ERA, batting average — can move it, inside a tight band.
Team totals are not tagged. Need at least 50% and -120 or longer.
Prefer the 57–72% band. Skip 0.5-line dart throws. One prop per game, locked, graded off the box.

Grok Expert (grok-4.6, high reasoning) can overlay the displayed moneyline winner from the same ratings, records, and DraftKings market. If Expert is down, the local model stands in and the card says so. The spread/total book and the player-prop book are always the local desk — they grade the frozen card, not a later rebuild.

What we never do

Invent a spread, total, or player price when DraftKings has not posted one.
Flip a locked card after kickoff, or after a final, or because a later desk liked someone else.
Grade a different player, line, or side than the one that was posted.
Lay shorter than -120 on either book.
Count the moneyline winner as the graded computer pick.

League constants

These are the knobs behind every projection. Home is the home-field add. σ is typical scoring noise. Shrink is how many games before a team’s own stats dominate the average.

League	Home	σ	Typical total	Shrink	Engine
NFL	+2.4 pts	13.6	44.8 pts	6 gms	Bell curve
NCAAF	+3 pts	16.2	53.5 pts	5 gms	Bell curve
NBA	+2.8 pts	12	224 pts	10 gms	Bell curve
WNBA	+2.2 pts	10.5	162 pts	8 gms	Bell curve
NCAAB	+3.4 pts	11.5	144 pts	8 gms	Bell curve
MLB	+0.14 runs	3.1	8.7 runs	22 gms	Poisson counts
NHL	+0.22 goals	1.7	6.1 goals	12 gms	Poisson counts
MLS	+0.28 goals	1.35	2.85 goals	8 gms	Poisson counts
EPL	+0.32 goals	1.28	2.75 goals	8 gms	Poisson counts
La Liga	+0.34 goals	1.26	2.55 goals	8 gms	Poisson counts
Bundesliga	+0.3 goals	1.32	3.05 goals	8 gms	Poisson counts
Serie A	+0.3 goals	1.24	2.6 goals	8 gms	Poisson counts
Ligue 1	+0.33 goals	1.3	2.7 goals	8 gms	Poisson counts
Champions League	+0.26 goals	1.34	2.9 goals	6 gms	Poisson counts
Nations	+0.22 goals	1.36	2.7 goals	4 gms	Poisson counts

UFC, ATP, WTA, F1, PGA skip a predicted score and use win probability only.

The stats, in one list

PF / PA: Season points (or goals/runs) scored and allowed. Converted to a per-game pace, then compared to league average.
Win%: Wins plus half of ties, over games played. Nudges attack and defense so winning teams rate a bit better than their raw scoring.
Form: Last six results, recent games weighted more. A small add-on — not a second model.
FPI / BPI: ESPN’s power index, in points of expected margin. Stitches preseason talent into week one.
Implied DK score: From the posted spread and total. Home team implied points = (total + home margin) / 2.
σ (sigma): How messy a sport’s scores are. Larger σ means a 3-point edge is closer to a coin flip.
Cover%: Share of the margin curve on the right side of the DK spread. 50% is a pick’em versus the number.
Over%: Share of the total curve above the DK total. Same idea, wider noise.
Juice: American odds on the posted side. -120 is the shortest price the computer will lay. −110 is a typical player-prop number.
Lock: The first posted spread/total and the first tagged player prop are frozen. Finals grade those cards against the box, not a later desk.

Want to see the ratings without a market prior? Run a custom matchup. Open any game for the live factor list on that card.