How Sportsbooks Control the Betting Market, and What Our Research Found

By Full Spectrum Sports Intelligence · 2026-10-09

Sportsbooks don't need to predict games better than bettors. They set the price, take a margin on every leg, and limit customers who keep beating them. Our research, tested in the open.

1. Thesis statement

Sportsbooks don't need to predict games better than bettors. They set the price, take a built-in margin on every leg, take a larger one on same-game parlays, and limit customers who keep beating them. Any edge that survives that has to be measured honestly, against a realistic book, on days the model never saw. Our main finding is uncomfortable but important: most "edges" in sports betting vanish once you test them against a book that prices as well as you do. Full Spectrum Sports Intelligence exists to run that test in the open, and to say which signals hold up and which don't.

2. How sportsbooks control the market: hold/vig, SGP correlation pricing, limits, line moves

Hold and vig on a single bet

Sportsbooks do not need to "beat" you on every ticket. They build a margin into the price of every market. That margin is usually called vig (the juice on the odds) or hold (the share of handle the book keeps after bets settle).

A standard two-way market priced -110 / -110 is the clearest example. Each side's implied probability is 110 / (110 + 100) = 52.38%. Adding both sides gives 104.76%, so about 4.5% of every dollar wagered is built in for the house before anyone wins or loses. Over a large sample, that structural take is the book's business model—not a lucky run of scores.

State regulators publish the aggregate version of the same idea. In New Jersey, the Division of Gaming Enforcement's December 2024 year-end report put completed-events sports wagering win at 8.4% of handle overall. Broken out by category, parlays kept 17.5% of $3.67 billion in parlay handle—while football, basketball, and baseball single-sport categories held between 2.9% and 4.2%. Through November 2025, the same NJ series showed parlays at an 18.7% win percentage on $3.57 billion of parlay handle, against a statewide completed-events win of 9.3% (NJ DGE December 2024; NJ DGE November 2025).

That product mix is the public record: multi-leg tickets carry a much larger house share than straight bets.

Why parlays compound the house edge

A traditional (cross-game) parlay multiplies leg probabilities as if the legs were independent. Even without correlation tricks, vig compounds. Two fair coin-flip markets at -110 each have a joint fair probability near 25%, but the book is already charging roughly 4.5% on each leg. Stack three or four legs and the bettor is paying that structural take repeatedly while chasing a larger payout.

Same-game parlays (SGPs) go further. Outcomes inside one game are not independent. A quarterback's passing-yards over, that team's win, and the game total over tend to rise and fall together. Multiplying the posted single-leg odds as if they were separate games understates the true chance that all of those legs hit. A book that paid "independent" parlay prices on positively correlated same-game legs would be giving the customer free money.

How books price SGP correlation

Public math explainers (including Wizard of Odds on SGP correlation) describe the industry pipeline in plain terms:

1. Estimate each leg's marginal (standalone) probability from the book's model and market.
2. Estimate the joint probability with a correlation model—empirical co-occurrence from past games, conditional probabilities, or methods such as Gaussian copulas that keep the marginals while adding a dependence structure.
3. Convert that joint probability into fair odds, then add margin.

When legs are positively correlated, the joint probability is higher than independence implies, so the fair payout is shorter. The book then applies vig on top of that correlation-adjusted fair price. When legs are negatively correlated, the joint probability falls and a fair book would lengthen the payout—but recreational tickets cluster on the narrative, positively correlated combos, which is exactly where the pricing adjustment is largest.

Worked intuition from that literature: a three-leg same-game ticket that looks like roughly 14–16% under independence can sit near 19–21% once realistic positive correlation is included. Books therefore offer a shorter SGP price and keep a house edge that is often several times a single-bet vig. The customer sees one payout—not a line item splitting "correlation tax" from "leg vig."

Why SGPs are among books' highest-margin products

Regulator data and operator disclosures point the same way.

• Regulators: New Jersey's completed-events tables (linked above) show parlays holding in the high teens while major single-sport categories hold in the low-to-mid single digits. That gap is structural, not a one-month fluke.
• Operators: Flutter Entertainment (FanDuel's parent) told investors that U.S. sportsbook Q4 2025 structural revenue margin reached 15.5%, up 90 basis points year over year, "driven by continued increase in parlay penetration" (80 bps of handle mix), with full-year 2025 structural margin at 14.2% (Flutter Q4 2025 earnings release). In other words, the company itself ties structural hold expansion to more parlays as a share of handle—not merely to lucky scores.

Three forces reinforce that margin: information asymmetry (joint models most customers never see), product design (storytelling and lottery-style upside), and selection (the "feel smart" stacks—win + star over + game over—are the positively correlated ones books haircut best).

None of this means every SGP is "rigged." It means the product is engineered so the house's edge is larger, and more durable, than on a carefully priced moneyline. Independent analytics matter when they measure that edge against a book that already knows the correlations—not against a calculator that pretends legs are independent.

Limits and line moves

The margin is built into every price. We took the closing DraftKings prices that ESPN publishes for every 2026 game we could get, converted both sides to implied probabilities, and added them up. The amount over 100% is the book's hold. The median closing hold on a two-way moneyline was 4.58% in MLB (n=2,454 games), 4.27% in the NFL (n=65, weeks 1–5), 4.20% in college football (n=323), 4.25% in the WNBA (n=351) and 4.28% in the NHL (n=57, season to date). Totals were similar: 4.34%–4.75%. A bettor has to win more than half their bets just to break even, before any skill shows up. (Source: DraftKings open and close prices published by ESPN's public odds API, https://sports.core.api.espn.com.)

Winners get limited. A Massachusetts Gaming Commission data analysis presented Sept. 30, 2025 found that 0.64% of the state's sportsbook accounts were limited. Of those, 57.6% were cut to 1–24% of the default maximum bet. It also found that "players who consistently beat the closing line are more likely to have a lower stake factor," while losing players were more likely to have their limits raised and to be made VIPs. The book doesn't have to out-predict a sharp bettor; it can simply stop taking their action. (Sources: MGC meeting minutes, https://massgaming.com/wp-content/uploads/Meeting-Minutes-9.30.25-OPEN.pdf; Covers, https://www.covers.com/industry/massachusetts-limiting-sports-betting-prediction-markets-exchanges-sharps-box-september-2025; Boston Herald, https://www.bostonherald.com/2025/09/30/winning-gamblers-face-betting-limits-losers-made-vips-gaming-commission-learns/.)

Lines move, and the closing line is the benchmark. From open to close, the total changed in 84.6% of NFL games (mean move 2.20 points, n=65), 84.4% of college football games (2.10 points, n=390) and 86.9% of WNBA games (2.26 points, n=351). MLB totals moved in 56.4% of games (0.34 runs, n=2,454). The point spread moved in 81.5% of NFL games but only 10.8% of MLB run lines, where the price moves instead of the line. Books use that movement to find out who is right. As the Massachusetts data shows, the customers who keep getting a better number than the close are the ones who get limited. (Source: DraftKings open and close prices published by ESPN's public odds API, https://sports.core.api.espn.com.)

3. What our research found: backtests, signals, ATS/cross-book findings

These results come from Full Spectrum's own backtest (Oct 9, 2026). We describe them in aggregate here; the detailed return figures are not yet published.

How we tested.

• We replayed the 2026 MLB season (42,627 settled player-prop legs, July 1 to Oct 8) using only information available before first pitch.
• We picked game days at random and held a final block of days back untouched.
• We priced every leg against two kinds of synthetic book:
• a "soft" book that knows less than our model;
• a "sharp" book that already prices with everything our model knows.
• Same-game parlays were charged a hold that grows with leg count: 9% for 2 legs, 14% for 3 and 22% for 4. These are our modelling assumptions, not quotes from any book.
• We did the same, more cheaply, for NHL and WNBA.

What we found (aggregate only).

• Against the soft book, our best MLB design showed a positive return in the season-long cross-validation.
• Against the sharp book, that edge disappeared.
• On the untouched final 10 days it lost money.
• No sport reached an 80% pick hit rate at break-even or better on held-out days. The one small pocket above 80% was mostly heavy favourites, whose low payouts erase the hit rate.
• More realistic correlation assumptions for same-team hitters and pitcher-vs-hitter legs priced slips better than the old ones, and never did worse.

Why that matters. If a model that we built and tuned can't beat a book that knows what it knows, a casual bettor's "lock" usually can't either. The edge, where one exists, sits in where a book is slow or careless, not in raw prediction. Section 4 covers those places.

4. Where books misprice, and why independent analytics matter

Sportsbooks are very good at pricing liquid, high-attention markets. They are less perfect everywhere else. Mispricing is not a conspiracy; it is a resource problem. Trading desks and models concentrate on moneylines, spreads, and totals that take the most handle. Everything else—long prop menus, low-profile players, late scratches, weather, and books that update slowly—gets less attention. That is where independent research has something useful to say.

Correlation mistakes and early SGP products

SGP pricing requires a joint model. When that model is wrong, the error shows up as a payout that is too generous or too stingy relative to the true co-occurrence of legs. Early SGP products, and thinner books copying SGP menus without deep historical matrices, historically struggled most on unusual combinations and on negative-correlation constructions that recreational bettors rarely build (Wizard of Odds; Wager Theorem on SGP pricing).

The practical takeaway for an authority site is not "chase every correlated stack." It is: compare offered SGP prices to a defensible joint probability, not to an independence calculator. A short correlated payout can still be fair; a long one can still be a trap if the book's margin after correlation remains large.

Slow books, stale lines, and cross-book gaps

Not every sportsbook updates at the same speed. When injury news, lineup confirmation, or sharp money hits a consensus line, slower books can briefly post yesterday's price. Public explainers of stale-line detection emphasize peer disagreement, open-versus-suspended status, and commence-time matching before treating an outlier as real edge (Parlay API on stale lines).

Player props widen that gap. A July 2026 cross-book study of 2,763 MLB and WNBA prop points offered at the same line by at least two of five books near 90 minutes before start found a median gap of 2.27 percentage points of implied probability on overs (90th percentile 5.20) and 1.59 points on unders (90th percentile 3.74) (PropsBot line-shopping study). That is not a guaranteed profit study—it measures price dispersion, not closing-line value—but it documents that "the market" is often several markets at once.

Low-limit and low-profile props compound the issue: less sharp money, slower human review, and models that share parameters across players who do not share roles. News latency (a starter ruled out, a bullpen game, a weather delay) often moves the main total first and the dependent props later. Wind, precipitation, and temperature affect passing games and totals; books that lean on paper openers without refreshing weather can leave totals and pass-yard markets behind the physical conditions of the kickoff.

What our own research found (aggregate, with caveats)

Full Spectrum's Oct 9, 2026 backtests are blunt about limits:

• We replayed a large 2026 MLB player-prop sample with pre-game information only, held days out of tuning, and priced slips against a soft synthetic book (knows less than our model) and a sharp synthetic book (already prices with our best features).
• The best MLB design showed a positive return against soft, model-like pricing in full-season cross-validation.
• Against sharp pricing, that edge disappeared, with a confidence interval that spans zero.
• The final untouched block of days was negative even against the soft book.
• Discrimination within markets was weak on the accuracy audit (AUCs only modestly above 0.5 for major prop families), and reported "skill" can look better than it is when markets with different base rates are pooled.

The honest reading matches Section 3: most apparent edges are an artifact of who you are pricing against. A stack that beats a naive Poisson or last-five average is not the same as a stack that beats a sharp desk. Until real multi-book closing prices are in the dataset, any ROI figure is a hypothesis about book softness—not a claim of risk-free profit.

Why independent analytics still matter

If books already captured every inefficiency, regulator hold percentages would not differ so sharply between parlays and singles, operators would not cite parlay mix as a structural-margin driver, and cross-book prop prices would not diverge by multiple probability points near first pitch. The remaining gaps are operational:

• Where the book is slow (props, news, weather, secondary books).
• How SGPs are joint-priced versus sold.
• Whether a claimed model edge survives a sharp counterparty and holdout days.

Full Spectrum Sports Intelligence's role, in the Unusual Whales sense, is to publish that measurement: methods, confidence intervals, and failures included—not a tip sheet of tonight's legs. Mispricing exists; proving it is harder than finding a narrative that fits the box score.

5. Data and methodology

Sources. Free, public data only. No logins, no paid feeds, nothing scraped against a site's robots.txt.

• Prices: ESPN's public odds API (https://sports.core.api.espn.com), which publishes DraftKings open and close prices plus player-prop and alternate-line ladders. Kalshi's public market API (https://api.elections.kalshi.com/trade-api/v2). Polymarket's public Gamma API (https://gamma-api.polymarket.com).
• Results, lineups and umpires: MLB Stats API (https://statsapi.mlb.com) and ESPN scoreboards.
• Weather: Open-Meteo's archived model forecasts (https://open-meteo.com).

What we built (2026 season, through Oct 8).

• Open/close lines for 3,348 games: MLB 2,483, college football 390 (FBS weeks 1–5), WNBA 353, NFL 65 (weeks 1–5), NHL 57 (season to date).
• 6,682 team-game against-the-spread rows, tagged favorite/underdog, home/away, rest days, division game and primetime.
• 2,483 MLB games with the home-plate umpire and full-game strikeouts, walks and plate appearances (league K rate 22.17%). Of these, 2,475 are joined to the closing total: 1,180 overs, 1,182 unders, 113 pushes.
• 4,966 starting-pitcher rows comparing the announced probable with the actual starter.

How we keep it honest.

• Every number is computed from the public sources listed above. Anything we couldn't get from a free source is marked "unavailable", never estimated.
• What we could not get:
• A sportsbook's actual same-game-parlay price: no free public source quotes it.
• Historical public betting percentages: the main free splits page disallows automated collection in its robots.txt (https://data.vsin.com/robots.txt).
• A weather forecast issued exactly three hours before first pitch: we use the closest archived model run and a 24-hour-earlier run, and label both.
• A bug we caught: our first MLB join matched some games to the previous day's game in the same series. We re-matched every game by start time and re-ran everything downstream.

Going forward. The app now records the real book price of every leg at the moment a user builds a slip, including strikeout main and alternate lines. It also snapshots each game at open, three hours out, one hour out and at close. That lets us measure closing-line value on our own selections, the same yardstick the Massachusetts regulators found books use to decide who gets limited. We need at least 30 days of that forward data before we will publish any result from it.

6. Conclusion and what Full Spectrum Sports Intelligence offers

Sources

• NJ DGE December 2024 gaming revenue results — completed-events win % by category; parlays 17.5% YTD 2024
• NJ DGE November 2025 gaming revenue results — parlays 18.7% YTD through Nov 2025
• Flutter Entertainment Q4 2025 earnings release — US structural revenue margin 15.5% Q4 / 14.2% FY; parlay penetration
• Wizard of Odds — Same-Game Parlays: The Mathematics of Correlation
• Wager Theorem — How sportsbooks price same-game parlays
• PropsBot — Player prop line-shopping study (July 2026)
• Parlay API — Detecting stale sportsbook lines