Quant Memo
Foundational

Craps: Working Out the House Edge

Deriving the house edge on a pass-line craps bet from scratch, using conditional probability across the come-out roll and the point phase — the standard 'work out this casino game's edge' interview question.

Prerequisites: Expected Maximum of Several Dice

Question. In craps, a "pass line" bet wins if the first roll (the come-out roll) is 7 or 11, loses if it's 2, 3, or 12, and otherwise the rolled number becomes "the point" — the shooter keeps rolling until either the point comes up again (bet wins) or a 7 comes up first (bet loses). What's the house edge on this bet? This is a genuinely multi-stage conditional-probability problem, and it's a favorite because it forces you to combine several probabilities correctly rather than just compute one.

Stage 1: the come-out roll

With two dice, sums have these counts out of 36: 7 has 6 ways, 11 has 2 ways, so P(win on come-out)=8/36P(\text{win on come-out}) = 8/36. Sums 2, 3, 12 have 1+2+1=41+2+1=4 ways, so P(lose on come-out)=4/36P(\text{lose on come-out}) = 4/36. The remaining 24/3624/36 of rolls establish a point of 4, 5, 6, 8, 9, or 10.

Stage 2: given a point is established, what's the win probability?

If the point is pp, the shooter keeps rolling until pp or 7 appears, ignoring all other numbers (they're just re-rolled). Given that only pp or 7 matters, the win probability is

P(winpoint=p)=ways to roll pways to roll p+ways to roll 7.P(\text{win} \mid \text{point}=p) = \frac{\text{ways to roll } p}{\text{ways to roll } p + \text{ways to roll } 7} .

For point 4 (3 ways) vs. 7 (6 ways): P=3/9=1/3P = 3/9 = 1/3. Point 5 (4 ways): P=4/10=2/5P = 4/10 = 2/5. Point 6 (5 ways): P=5/11P = 5/11. By symmetry, points 8, 9, 10 mirror 6, 5, 4 respectively.

Worked example: assembling the total win probability

Total win probability sums the come-out win, plus for each point value, the probability of establishing that point times the conditional win probability given that point:

P(win)=836+pP(point=p)P(winpoint=p).P(\text{win}) = \frac{8}{36} + \sum_{p} P(\text{point}=p) \cdot P(\text{win}\mid \text{point}=p) .

Plugging in each point (4 and 10 each have 3/36 chance to be established, with win prob 1/3; 5 and 9 each 4/36 with win prob 2/5; 6 and 8 each 5/36 with win prob 5/11):

P(win)=836+2(33613)+2(43625)+2(536511).P(\text{win}) = \frac{8}{36} + 2\left(\frac{3}{36}\cdot\frac{1}{3}\right) + 2\left(\frac{4}{36}\cdot\frac{2}{5}\right) + 2\left(\frac{5}{36}\cdot\frac{5}{11}\right) .

Computing each term: 8360.2222\frac{8}{36} \approx 0.2222; 2×336×13=2360.05562 \times \frac{3}{36}\times\frac{1}{3} = \frac{2}{36} \approx 0.0556; 2×436×25=3.2360.08892 \times \frac{4}{36}\times\frac{2}{5} = \frac{3.2}{36} \approx 0.0889; 2×536×511=4.545360.12632 \times \frac{5}{36}\times\frac{5}{11} = \frac{4.545}{36} \approx 0.1263. Summing: P(win)0.2222+0.0556+0.0889+0.12630.4929P(\text{win}) \approx 0.2222+0.0556+0.0889+0.1263 \approx 0.4929, or about 49.29%49.29\%.

Since this is an even-money bet, the house edge is

edge=12×P(win)12(0.4929)=0.0141,\text{edge} = 1 - 2 \times P(\text{win}) \approx 1 - 2(0.4929) = 0.0141 ,

about 1.41% — matching the well-known real-world figure for the craps pass line, one of the best bets on a casino floor precisely because the multi-stage structure keeps the edge small compared to, say, roulette's 5.26%.

P(win | point) 4 (1/3) 5 (2/5) 6 (5/11) 8 (5/11) 9 (2/5)
Conditional win probability rises the closer the point is to 7, since 6 and 8 have the most ways to roll relative to 7's six ways — points 4 and 10 (not shown, mirrors 4) are the hardest to win.

Multi-stage betting games are solved by conditioning on the intermediate state (here, "what's the point") and combining stage probabilities with the law of total probability: P(win)=statesP(state)P(winstate)P(\text{win}) = \sum_{\text{states}} P(\text{state}) \cdot P(\text{win} \mid \text{state}). The house edge follows immediately once you have the true win probability and the payout odds.

The most common error is forgetting that once a point is established, rolls that are neither the point nor 7 simply don't count — treating this as "P(point before 7) over all 36 outcomes per roll" without conditioning down to just the two relevant outcomes gives the wrong per-roll probability, even though the final win probability formula above handles it correctly by construction.

Related concepts

Practice in interviews

Further reading

  • Epstein, The Theory of Gambling and Statistical Logic, ch. 5
ShareTwitterLinkedIn