ABC Tools

League Season Simulator, Explained

After reading this you will understand how a Monte Carlo season model turns a partial league table into title, top-N and relegation odds, how to reproduce those numbers by hand for one match, and where the model quietly lies to you.

What the simulator does

A league table on a given day is a snapshot, not a verdict. Arsenal leading Liverpool by one point with one game in hand tells you Arsenal are likely champions, but "likely" is a number, and the table alone does not give it to you. The League Season Simulator supplies that number by playing the rest of the season thousands of times.

Each simulated season assigns every remaining match a random result drawn from win, draw and loss probabilities based on how strong each team has looked so far. Play all remaining games, sort the final table, and record who finished first, who finished in the top four, and who finished in the bottom three. Do that 5000 times and count. If Arsenal win the title in 3100 of 5000 runs, their title probability is 3100 / 5000 = 0.62, or 62 percent.

Nothing here is exact. The output is a distribution over outcomes, and its width honestly reflects how much football is left to play.

When to use it, and when not to

Use it when you have a current table and want odds rather than a hunch. It shines mid-season, when there are enough games played to estimate strength but enough remaining to leave the result in doubt. With 24 of 38 games gone, roughly 63 percent of the season is decided and 37 percent is still random, which is the interesting zone.

Be cautious in two situations. Very early, after two or three games, points per game is a noisy estimate of strength, so treat the probabilities as vague. Very late, with one or two games left, the model ignores the actual fixture list, so a team that must still play the two strongest sides is treated the same as one facing the two weakest. For a title race decided on the final day, trust the fixtures over this model.

This model has every team play "the average of the league" for its remaining games. There is no schedule, no head-to-head, and no home advantage. Two teams level on points get identical remaining difficulty, even if their real run-ins differ sharply. Read tight races as approximate.

The strength model and match probabilities

Each team's strength is its smoothed points per game. Raw points per game is simply points divided by games played:

\text{ppg}_i = \frac{\text{points}_i}{\text{played}_i}

Here points_i and played_i are the values you paste for team i. Arsenal on 55 points from 24 games have 55 / 24 \approx 2.29 points per game. Smoothing pulls extreme early-season rates toward the league average of about 1.35 per game, which is why a team on a small sample gets sensibly wider odds.

For a single match between team i and team j, the model compares their strengths. A convenient way to turn a strength gap into a win probability is a logistic curve:

P(i \text{ beats } j) = \frac{1}{1 + e^{-k (\text{ppg}_i - \text{ppg}_j)}}

The constant k sets how much a strength gap matters, and the draw share is carved out around the middle so that win, draw and loss sum to 1. The exact scaling is calibrated so that a full season of these games reproduces realistic points totals. The key intuition: equal teams sit near a coin flip adjusted for draws, and a large ppg gap pushes the stronger team toward a high win probability without ever reaching certainty.

Worked example with the demo table

Arsenal versus Newcastle, one match

Load the demo table (Arsenal 55/24, Liverpool 54/23, down to Brighton 35/23) and consider one remaining fixture: Arsenal against Newcastle (38 from 24).

  1. Arsenal ppg: 55 / 24 = 2.292. Newcastle ppg: 38 / 24 = 1.583. Gap: 0.708.
  2. Apply the logistic with a calibrated k \approx 1.1: exponent is -1.1 \times 0.708 = -0.779, so the raw favouring factor is 1 / (1 + e^{-0.779}) = 0.685.
  3. Reserve a draw share of about 0.26 for a gap this size, then split the rest. Arsenal win probability lands near 0.55, draw near 0.26, Newcastle win near 0.19.
  4. Draw one uniform random number in [0, 1). Below 0.55 is an Arsenal win (3 points), 0.55 to 0.81 a draw (1 point each), above 0.81 a Newcastle win.

Repeat that for every unplayed fixture of every team, sum the points onto the current totals, and you have one simulated final table. The whole season simulator is this single-match step run across the schedule and then across 5000 seasons.

Expected points from one match is easy to check: 3 \times 0.55 + 1 \times 0.26 + 0 \times 0.19 = 1.91 for Arsenal, and 3 \times 0.19 + 1 \times 0.26 = 0.83 for Newcastle. Averaged over 14 remaining games, small per-match differences accumulate into the expected final points column.

Reading the output

The simulator reports four things per team: title odds, top-N odds, bottom-N odds, and expected final points. Treat each as a frequency out of your iteration count. With 5000 runs, a reported 62 percent means 3100 seasons ended with that team first.

Expected final points is the average of the final total across all runs, not a prediction of the exact number. If Arsenal average 84.2 points, individual runs might range from 78 to 91. The spread matters as much as the average.

Title odds are concentrated among the top three, while top-four odds stay high for the leaders and fall off around Villa and Spurs. Your exact numbers will vary with iterations and calibration.

Notice that title percentages across all teams add to about 100, because exactly one team wins. Top-four percentages add to about 400, because four teams finish there. Use that as a sanity check on any table of results.

Monte Carlo error: how many runs you need

Because each odds figure is a proportion estimated from a sample of seasons, it carries sampling error. For a probability p estimated from n runs, the standard error is:

\text{SE} = \sqrt{\frac{p(1 - p)}{n}}

At p = 0.5 and n = 5000, the standard error is \sqrt{0.25 / 5000} = 0.00707, about 0.7 percentage points. So a reported 50 percent is really 50 plus or minus roughly 1.4 points at two standard errors. Quadruple the runs to 20000 and the error halves to 0.35 points, because error shrinks with the square root of n: to cut it in half you multiply runs by four.

At 5000 iterations a 50 percent estimate has a standard error near 0.7 points; at 1000 iterations it grows to about 1.6 points; at 50000 it shrinks to about 0.22 points. Fewer runs mean noisier odds.

Common mistakes

The most frequent error is over-reading a tight race. If two teams sit within a point and the simulator says 52 to 48, remember that gap is inside the Monte Carlo error at low iteration counts and inside the model error always. Do not report it as a firm edge.

A second mistake is trusting a very early table. After three games a team on nine points shows 3.0 ppg, which no side sustains. Smoothing helps, but the honest reading of an early simulation is "wide," not "confident."

A third is ignoring the missing fixture list near the end. On the final weekend, who plays whom decides everything, and this model averages that away. A fourth is mismatching total_games to your league: if you set 38 for a 34-game league, every team gets four phantom fixtures and the odds drift toward the strong teams.

Related tools

The single-match engine here is deliberately simple. If you want to model matches from a rating gap directly, the Elo Rating Calculator converts two ratings into an expected score. To model scorelines and win-draw-loss from expected goals, use the Soccer Score Simulator (Poisson). For knockout formats rather than a league, the Playoff Series Probability Calculator handles best-of-N series, and the Penalty Shootout Probability Calculator covers the tie-break. For a fully worked serve-to-match chain in a different sport, see the Tennis Match Probability Calculator.

Frequently asked questions

Why do my odds change slightly every time I run it?

Each run uses fresh random numbers, so results wobble within the Monte Carlo error. At 5000 iterations a 50 percent figure moves by roughly a point run to run. Raise iterations to steady it, or accept the wobble as the honest uncertainty.

How many simulations should I use?

5000 is a good default for reading odds to the nearest few percent. Use 20000 or more if you care about differences smaller than one percentage point, remembering the error falls only with the square root of the count.

Why does a team with fewer games sometimes get lower odds than expected?

Fewer games played means more matches remaining, so more randomness sits between now and the finish. A narrow points lead with many games left is less secure than the same lead with few games left.

Does the model include home advantage or specific fixtures?

No. Every team plays the league average for its remaining games. Head-to-heads, venue and schedule difficulty are not modelled, which is why the footnotes call tight races approximate.

Can I use this for basketball or another league?

Yes, if wins, draws and losses (or wins and losses) map onto a points system and you set total_games correctly. The strength model is sport-agnostic; only the points per game input changes.

Try it yourself in the tool

Related tools

Elo Rating Calculator

Expected score and rating change for any two Elo ratings and K-factor.

Playoff Series Probability Calculator

Chance of winning a best-of-3/5/7 series from your single-game win probability.

Soccer Score Simulator (Poisson)

Win/draw/loss and most likely scorelines from two expected-goals numbers.

Badminton Match Probability Calculator

From serve-rally win % to game and match probability with the 30-point cap.

Squash Match Probability Calculator

From serve-rally win % to game and match probability with PAR-11 scoring.

Table Tennis Match Probability Calculator

From serve-point win % to game, deuce and match probability for best-of-3/5/7.