How to Build a +EV Lottery Ticket on NBA Threes

To test a longshot NBA three-point ticket for positive expected value, model made threes from attempts and percentage, simulate the joint outcome of every leg rather than multiplying them, convert that probability into fair odds, and compare with the price offered. Treat large apparent edges as probable model error and size stakes for extreme variance.

What the video proposes

The short from the NoahWinsBets channel, posted on 18 November 2026, makes one argument: a longshot same-game parlay can still carry positive expected value if its legs are negatively correlated and the sportsbook misprices that correlation. Its main example pairs a player's Over on made three-pointers with the same player's Under on total points. A player who hits five threes and almost nothing else finishes on 15 points.

The video says that with line shopping it has seen this kind of combination priced as high as +18000, and that at that price the outcome only needs to land about once every 180 games to break even. It also suggests the mirror image for players with strong mid-range games: Under on threes, Over on points. Its closing advice is to size these as small, high-variance positions. Those are the video author's claims; what follows is a method for testing them, with illustrative numbers.

Step 1: model the distribution of made threes

Start with the single leg. Made threes are attempts times a hit rate, so the natural model is binomial: n attempts, each made with probability p. With 8 attempts at 38 percent, the binomial gives about 36 percent for four or more makes, 14.4 percent for five or more and 3.9 percent for six or more.

Attempts are not fixed, though. They depend on minutes, game script and role, so a Poisson model on makes, with the mean set to expected attempts times percentage, is a common shortcut. It has fatter tails than the fixed-attempt binomial: at a mean of 3.04 makes, Poisson puts five or more at about 19 percent and six or more at 8.8 percent, more than double the binomial figure for six. Longshot prices live in that tail. Better: model attempts as uncertain, driven by minutes, and makes as binomial given attempts, then check the tail against the player's game logs.

Converting a probability to fair odds

Fair decimal odds are 1 divided by the probability. A 14.4 percent leg is fair at about 6.94, or +594 in American terms. A 0.82 percent joint outcome is fair at about 122, or +12100. A price is only worth considering when the decimal odds on offer exceed your fair odds by more than your uncertainty about the probability.

Step 2: model the joint outcome, not the legs

Multiplying leg probabilities assumes independence, and these legs are nowhere near independent. Two forces pull in opposite directions. Minutes create positive correlation: more floor time means more threes and more points, which makes Over threes with Under points less likely. Shot mix creates the negative relationship the video relies on: for a given number of made threes, the points total depends on what else the player does.

The practical tool is simulation. Draw minutes from a distribution, draw three-point, two-point and free-throw attempts from per-minute rates, draw makes from binomials, add up points, and count how often both legs land. An illustrative player at about 30 minutes, 8.1 three-point attempts at 38 percent, 3 two-point attempts at 50 percent and a little over one free throw attempt per game produced, over 3 million simulated games, these figures for five or more threes with Under 15.5 points: about 20 percent for the threes leg, about 67 percent for the points leg, a naive product of 13.5 percent, and a simulated joint probability of 0.82 percent. Independence overstates the joint chance by a factor of more than 16. Fair odds in this toy model are about 122 in decimal terms.

Step 3: compare with the price on offer

Suppose a book offers the combination at +18000, decimal 181, the top price the video mentions. The implied probability is about 0.55 percent. Against the toy model's 0.82 percent, expected value is 0.0082 times 181, minus one, or about plus 48 percent per unit staked. An edge that large is more likely a model error than a gift.

Rerun the simulation with different inputs and the conclusion moves. Dropping three-point accuracy to 35 percent lowers the joint probability to about 0.70 percent, still above the break-even figure. Raising the two-point attempt rate by 40 percent, to a player who takes about 4.2 twos a game, cuts the joint probability to about 0.46 percent, with fair odds near 216 and negative expected value at 181. The single most important input is not the three-point percentage but how much non-three scoring the player does. That matches the video's shot-profile point: any edge sits in player-specific role and shot-mix data.

Availability and market rules

Same-game parlay menus differ by book and change often. Since the 2026 to 27 season opened on 20 October, ESPN has reported that the NBA and its partner sportsbooks removed parlays that combine multiple unders on the same player, along with some other prop types. An Over and Under combination is not in the reported list, but what each book offers is its own decision.

Step 4: plan for the variance

A bet that lands 0.82 percent of the time loses almost always. With that probability, the chance of zero hits in 100 tickets is about 44 percent, in 200 tickets about 20 percent, and in 300 tickets about 9 percent. Even a genuinely positive position can go hundreds of tickets without a single return, and one hit dominates the whole record.

Stakes have to be tiny relative to bankroll, well below even fractional Kelly on the estimated edge. Results cannot validate the model on any realistic timescale, so check calibration on the single legs, which settle far more often. And model error dominates: one changed input moved this example from a large edge to a loss.

The data this method needs

Testing these tickets properly takes three things: per-game player box scores going back several seasons (attempts, makes and minutes by shot type), historical prop prices for each leg across books so you can see where the singles traded, and same-game parlay quotes captured at the time, since those are rarely archived. Parlay Ledger's API returns odds, historical odds, game results and player statistics across sportsbooks in one unified format, which covers the first two and lets you store your own snapshots for the third. The MCP server's backtest_strategy tool can then run a rule over historical data rather than over a handful of memorable hits.

Gambling carries real financial risk. Treat this as a modelling exercise, set limits before you bet, and use the responsible gambling tools licensed operators provide.

Frequently asked questions

Can a longshot parlay really have positive expected value?
Yes, in principle. Expected value depends on whether the payout exceeds the true probability, not on how likely the ticket is. A longshot is positive only if your joint probability estimate beats the implied probability by more than your model error. For correlated same-game legs that estimate is hard, so most apparent edges turn out to be modelling mistakes.
Should made threes be modelled with a binomial or a Poisson distribution?
Binomial fits when attempts are known: n shots at probability p. Poisson on makes is a simpler shortcut but gives fatter tails. Neither alone is ideal. A stronger model draws attempts from a distribution tied to minutes and role, then makes as binomial given attempts, and checks the tail against the player's historical games.
Why does Over threes with Under points pay so much?
Because both legs landing requires a narrow outcome: many made threes and almost no other scoring. The legs are negatively linked through points, and minutes add positive correlation that makes it rarer still. Whether the price is generous depends on the player's shot profile, especially how many twos and free throws they usually produce.
How many tickets would it take to know whether the edge is real?
Far more than most people will place. At about a 0.8 percent hit rate, there is roughly a 20 percent chance of no hits in 200 tickets even when the model is right. Results alone cannot settle it, so validate the inputs on single legs that settle often and compare your fair prices with sharp market prices.