The margin on a parlay is not the margin on a bet

MLB · 2–24 August 2026 · 447 games · 9 books · 9,670 two-sided market–book observations

A single MLB bet at these nine books carried an average theoretical hold of 5.01%. Combine three of those bets into a parlay and the hold is 14.30%. At five legs it is 22.67%.

You are not paying the margin once and spreading it across the legs. You are paying it on every leg, on the same stake. The margins multiply, because that is exactly how a parlay is priced — the book multiplies the decimal odds together, and the margin baked into each one comes along.

Most people know the vig on a standard bet is about 4.5%. Far fewer have looked at what it becomes at five legs. For comparison, the house edge on American roulette is 5.26% — so a single MLB bet here is roughly a roulette spin, and a three-leg parlay is close to three spins' worth of edge paid on one stake.

Percentage points between the cheapest and dearest book. The gap compounds exactly as the margin does.

This is the more defensible number, and the one you can act on. The all-books average in the board above leans toward whichever books we observed most — DraftKings is 1,694 of the 9,670 observations, lowvig 649. A gap between two books does not have that problem, because each book is measured against its own prices.

The obvious objection, tested

If the books quoted different games, the gap could be market mix rather than pricing. So here is the same comparison restricted to the 242 markets every one of the nine books quoted — identical games, identical market types, no mix to argue about:

The order barely moves and the gap survives: 2.58 points at one leg, 10.83 at five, on identical markets. Smaller than the full-sample figure, because 242 markets is a narrower and more liquid slice — but it is the like-for-like number, and it says the difference is pricing rather than which games each book happened to post.

Moneyline is the cheapest and the most variable — its interquartile range spans about three-quarters of a point, because moneyline margin moves with how lopsided the matchup is. Totals are the most consistent: the p25 and p75 sit five basis points apart, which is what a market priced off a standard template looks like.

Deliberately simple, so it can be checked rather than trusted.

1. Pin the extraction, then de-duplicate. The odds table is still being written to, so every figure here comes from observations strictly before 2026-08-24 20:37 UTC. Re-running against live data will not reproduce them; re-running with that cutoff will.
   
   A market polled more than once would otherwise be counted more than once, which inflates the apparent sample without adding information. So the unit of analysis is one observation per market per book — the most recent before the cutoff. That takes the raw price history down to 9,670 two-sided market–book observations, and 9,670 is the number any confidence in these figures should rest on.
   
   The cutoff moved once already. An earlier version of this page was pinned to 6 August and reported a 4.33% single-bet hold; three more weeks of collection put it at 5.01%. A pin makes figures reproducible, not permanent — so the date is stated rather than implied, and this one is final because collection has stopped.
2. Take two-sided markets only. Moneyline, spread and total each have exactly two outcomes. Every observation used here has both sides quoted by the same book at the same timestamp. The 1,806 market–book pairs where only one side was captured are dropped, not imputed — 16% of the sample. Worth naming rather than burying: books pull a side as kickoff approaches and this collection polls near kickoff, so the method and the exclusion are related. It is not a random loss.
3. Convert to implied probability and sum. For a market with decimal odds d₁ and d₂:

   S = 1/d₁ + 1/d₂

   A market with no margin sums to exactly 1.0. Everything above that is the book's overround.
4. Overround and hold are different numbers. This trips up a lot of published figures:

   overround        = S − 1
   theoretical hold = (S − 1) / S

   Overround is the margin expressed against a fair book. Hold is the share of total handle the book keeps on perfectly balanced action, and it is the smaller of the two. Both are in the board above so neither has to be taken on faith.
5. Parlay margin is the product, not the sum. A parlay's decimal price is the product of its legs, so its overround is the product of the legs' overrounds:

   S_parlay = S₁ × S₂ × … × Sₙ

   That step is arithmetic, not a model. The only assumption is which legs — the board uses the sample's average leg, which is why it is labelled as an average rather than as your parlay.
6. Sanity filter, which did nothing. Observations with S outside 1.0–1.5 would be excluded as malformed. Zero were excluded — every two-sided pair in the sample fell inside that range. The filter is reported because it was applied, not because it changed anything.

The caveats matter more than the headline, and leaving them out is how a study like this gets taken apart in the replies.

Every figure above comes from two files. Neither requires our database or our word.

game_id · market_id · bet_type_id · sportsbook_id · observed_at · selection_id · decimal_odds

The CSV is the sample after pinning and de-duplication — one observation per market per book, both sides, nothing else removed. Rebuilding the headline figure from it takes four lines in any language:

import csv, collections
rows = list(csv.DictReader(open('parlay-hold-observations.csv')))
mk, n = collections.defaultdict(float), collections.Counter()
for r in rows:
    k = (r['market_id'], r['sportsbook_id'])
    mk[k] += 1 / float(r['decimal_odds']); n[k] += 1
S = [s for k, s in mk.items() if n[k] == 2]
mean = sum(S) / len(S)
print(len(S), mean, (mean - 1) / mean)          # 9670  1.052770  0.0501

If that gives you something other than 9,670 observations and a 5.01% hold, we would like to know.

We are building a tool that shows what a parlay costs before you place it. This is the number the tool is built around, so it seemed reasonable to show our working rather than assert it.

It is worth being plain about what this finding is and is not. It will not make anyone a winning bettor. Nothing will — parlay betting has a negative expected return over time, and knowing the margin more precisely does not reverse that arithmetic. It only means you know what you are paying.

We take no money from sportsbooks. No affiliate links, no referral deals, no partnerships. The per-book table above is the reason that matters: nothing a book pays us could move a row in it.

The data, the method and the caveats are all above. If something here is wrong, we would rather hear it than not.