Simulating a season: how win probabilities are built
Playoff odds and championship probabilities come from simulating the schedule thousands of times. This article explains the method, the arithmetic of luck over 82 games and a seven-game series, and what a percentage means.
When a site tells you a team has a 78 percent chance of making the playoffs or a 12 percent chance of winning the title, it is not claiming to know the future. It is reporting the result of an experiment: the remaining season was played out on a computer many thousands of times, with the outcome of each game drawn at random from a probability, and the team made the playoffs in 78 percent of those simulated seasons. The method has a name, Monte Carlo simulation, and it is the standard way to turn a rating for each team into odds for every possible ending of a season.
The idea is easy to grasp and easy to misread. A probability is not a prediction. A 12 percent title chance does not say the team will fail; it says that in a world where the rating is right, about one in eight such teams wins. This article explains how the simulation is built, uses small calculations you can check yourself to show how much luck there is in an 82-game season and a best-of-seven series, and describes what separates a careful simulation from a naive one. It builds on the ratings discussed in Elo ratings in basketball: the idea and its uses and Net rating: the best single number for a team.
The ingredients
A season simulation needs four things.
- A rating for every team. It can come from an Elo system, a point-differential model, a player-based model or any other source. All that matters is that it can be converted into the probability that one team beats another.
- The schedule. Every game still to be played, with the home team identified, because home court matters, as shown in Home-court advantage in the numbers.
- The rules that turn results into outcomes. Seeding, tie-breakers, the play-in tournament, the structure of the playoff bracket and the format of each series. These are described in How the NBA playoffs work: seeding, series format and home court, How the NBA play-in tournament works and Tie-breakers in the NBA standings: how seeds are decided.
- A random number generator. The computer needs a way to decide each game, and the generator supplies it.
Step one: a probability for a single game
The first job is to convert two ratings into a win probability for one game. The details depend on the rating system. In an Elo system the gap between the ratings, plus a bonus for the home team, is fed into a logistic curve. In a point-differential system, the gap is converted into an expected margin, and the margin is converted into a probability using the fact that real results scatter around the expectation. A common working assumption is that a single game's margin varies from its expectation by something like a dozen points in either direction, which means a team expected to win by 5 has roughly a two-in-three chance to win, and a team expected to win by 10 has about an 80 percent chance. Those are approximations; the exact figures depend on the league and the era.
Whatever the method, the output is a number between 0 and 1 for each game. A game between a strong home team and a weak visitor might be 0.85. A game between two equal teams in a neutral building is 0.50. The simulation will ask the computer to treat each of these as a weighted coin flip.
Step two: play the season once
To simulate one run of the remaining schedule, the computer goes through the games one by one. For each, it draws a random number between 0 and 1. If the number is below the home team's win probability, the home team wins; otherwise the visitor does. A game with a 0.85 chance will go to the home team about 85 times in 100, but not every time, and in a single run any individual result can go either way.
After the last regular-season game, the computer has a complete set of results. It adds them to the games already played and builds the final standings, applying the tie-breaking rules for teams with the same record. Then it applies the postseason rules: who is in, who goes to the play-in, who gets which seed. It plays the play-in games and the playoff series the same way, drawing a random winner for each game using the probabilities for that matchup, and carries the bracket all the way to a champion.
One run produces one imaginary season. It might be a typical season or a strange one; the team the model likes least might win the title. That is fine. A single run means nothing by itself.
Step three: do it again, thousands of times
The simulation's power comes from repetition. Run the season 10,000 times, or 100,000, and tally the results. If a team made the playoffs in 7,800 of 10,000 runs, its playoff probability is 78 percent. If it won the title in 1,200 runs, its title probability is 12 percent. If it finished with 50 or more wins in 4,000 runs, there is a 40 percent chance of that. Any question about the shape of the season can be answered by counting.
More runs reduce the noise in the answer. With a few hundred runs, the estimate for a small probability bounces around; with tens of thousands it settles. For a probability near one half, the uncertainty from the number of runs alone drops to well under a percentage point with 10,000 runs. But the number of runs is not the main source of uncertainty. The main uncertainty is in the ratings, as described below.
How much luck is in a season?
A simulation makes the role of chance concrete, and you can see it with simple arithmetic. Suppose a team has a 60 percent chance to win each of its 82 games. The expected number of wins is 82 times 0.6, or about 49. But the actual total varies from year to year even if the team's true quality never changes. For a coin that lands heads with probability p over n tosses, the standard deviation of the total is the square root of n times p times one minus p. For 82 games at 0.6, that is the square root of about 19.7, or roughly 4.4 wins.
Roughly two-thirds of the time, then, such a team will finish within about 4 wins of 49 either way, from 45 to 53. And about one season in twenty it will finish more than 8 or 9 wins from its true level. A team that is genuinely a 49-win team will sometimes win 56 and sometimes win 42. For an average team with a 50 percent chance in every game, the standard deviation is about 4.5 wins. That is why teams so often improve or decline by a large number of wins from one year to the next without anything dramatic changing. It is the same phenomenon described in Regression to the mean: why extreme seasons rarely repeat, and it is why Pythagorean wins: using points to predict records is a useful check on a record.
The simulation captures that spread automatically, which is why its output is a distribution, not a single number. Instead of saying "this team will win 49 games," it says "this team will win between 45 and 53 about two-thirds of the time, and here is the chance of each total."
A tiny simulation by hand
You can feel the method with a toy case. Suppose two teams are tied for the last playoff place with three games left each, and every one of those games is a coin flip for both. Each team's remaining wins can be 0, 1, 2 or 3, with probabilities of one-eighth, three-eighths, three-eighths and one-eighth. Work out how often Team A ends up with more wins than Team B and you find it is about 34 percent, the same for Team B, and the remaining 31 percent or so of the time they finish tied. That is the entire logic of a simulation, written out in a few lines: every path is enumerated, or in a larger problem sampled, and the tally becomes the probability.
The tie matters, and it is where the rulebook enters. When the two finish level, a tie-breaker decides who gets the place, so each team's chance is 34 percent plus its share of the tied cases. Because the teams in this toy case are identical, the shares split evenly and each lands at 50 percent; in a real race, the head-to-head record and the other criteria tilt the split. A simulation of a full season does the same bookkeeping for all thirty teams at once, which is why it is built by computer and not by hand.
How much luck is in a series?
The same arithmetic applies to a playoff series. In a best-of-seven series, the first team to win four games advances. If the stronger team has a probability p of winning each game, the chance it wins the series is the sum of the ways it can win in four, five, six or seven games.
Take a team that is a 60 percent favorite in every game. Its chance to win the series works out to about 71 percent. A 70 percent favorite in each game wins the series about 87 percent of the time. Even a team that is an overwhelming 80 percent favorite in each game, a mismatch that would be extremely rare between two playoff teams, loses a series about one time in thirty. And a team that is an even 50-50 in each game wins the series exactly half the time, which is why a series between two equal teams is a coin flip.
The consequence is that a best-of-seven series filters out luck better than a single game does, but not by as much as many people assume. A 60 percent single-game favorite is a 71 percent series favorite, which means that upsets are not rare. They are the normal tax of the format. A simulation builds the bracket by chaining these series together: the chance to win a title is the product of the chances to win each round, taking into account who the opponent is likely to be. A team with a 71 percent chance in each of four rounds wins the title about a quarter of the time, since 0.71 multiplied by itself four times is about 0.25. In real brackets the first-round odds are higher and the later rounds lower, but the principle is the same: a title is the product of several uncertain events, and the product shrinks quickly.
This is also why the pre-season favorite in any sport wins the title less often than fans expect. In a field of 30 teams, being the most likely champion can mean a probability of 20 or 30 percent, which is a long way from a sure thing.
What separates a careful simulation from a naive one
The skeleton above is simple. The quality is in the details, and a few matter more than the others.
Rating uncertainty. A naive simulation treats each team's rating as known exactly. But a rating is an estimate, and there is uncertainty about whether a team is truly as good as it looks. A better simulation draws a slightly different "true" rating for each team in each run, in addition to the random game outcomes. The effect is to widen the range of results and to make extreme outcomes, both good and bad, somewhat more likely. Without this step, the simulation understates how often a team that looks only decent turns out to be a contender, and how often a leader turns out to be ordinary.
Changing ratings. Teams change during a season. Players are injured and return, trades happen, rotations shift. A careful simulation either updates the ratings as the simulated season goes on or adjusts them for known absences. This is the biggest weakness of simple rating methods, since the team that plays in April may not be the team that played in November.
Correlation. Results are not fully independent. If a team's star is hurt, all its remaining games are affected together. If a rating is too high, it is too high in every game. Treating every game as an independent coin flip understates the chance of extreme seasons. Drawing a team-level rating error in each run, as above, captures much of this.
Schedule details. Back-to-back games, long road trips and rest days affect results, as discussed in Back-to-backs, rest days and travel: how the schedule shapes strategy. Some models include them; many do not. The structure of the schedule itself, described in How the NBA schedule is made: 82 games, arenas and television, is an input that cannot be ignored when games remaining differ between teams.
Tie-breakers and format. The simulation must encode the league's rules exactly, including the play-in tournament and the order of tie-breakers. A mistake in the rules is a mistake in every run.
How to read a probability
A few habits keep you honest when you see a simulated number.
First, a probability is calibrated if events given 30 percent happen about 30 percent of the time. Over many forecasts, a good model is right about its probabilities even though it is wrong about many individual outcomes. A single event that the model rated at 10 percent happening is not evidence the model is broken; it should happen about one time in ten.
Second, a probability is only as good as the ratings behind it. A simulation does not add information. It converts the information in the ratings and the schedule into odds, so if the ratings are stale, the odds are too. Read a published probability as "given what the model knows."
Third, small changes in probability are not news. A move from 62 to 65 percent may be well within the noise of the method and the ratings. Look for changes that follow a real event, such as a long injury or a trade, and be skeptical of those that follow a single game.
Fourth, the tails are the part to remember. A team with a 3 percent chance of winning the title is not out of it, and a team with a 97 percent chance of making the playoffs can still miss. The remaining few percent is the share of the world the model considers possible.
What to take away
A season simulation is a counting exercise. Convert ratings to game probabilities, play the schedule with a weighted coin, apply the league's rules to get standings and a bracket, and repeat thousands of times. The result is not a forecast of one outcome but a map of how likely each outcome is. The arithmetic of an 82-game season, where luck alone moves a team's win total by four or five games in a typical year, and of a seven-game series, where a 60 percent favorite wins about 71 percent of the time, explains why upsets are ordinary and why even strong favorites win titles less often than fans expect. When you read a probability, ask what ratings produced it and how it handles uncertainty, and treat it as a statement about odds. You can see how the real season is unfolding in the standings and the playoff picture.