The PremPredict package helps you to generate sensible predictions for
individual games or an entire season of the Premier League.
You can find my automatically-updated Premier League predictor (that uses this codebase) on the landing page of its repo.
You can install the development version of PremPredict from GitHub with:
# install.packages("pak")
pak::pak("p0bs/PremPredict")I use a simplified version of David Firth’s approach and data from the Open Football repo on GitHub to predict the outcome of this season’s Premier League.
The predictions are based on a team’s strength, given its performance in recent times. But how should we define ‘recent’? In order to duck this question, you could choose a number of different time periods. Please also see some further disclaimers that are in these notes for my automatically-updated Premier League predictor (as linked above).
Here is an example analysis, using data collected towards the end of the 2025/26 season.
First, we collect, combine and tidy the results data.
library(PremPredict)
data("example_thisSeason")
results_combined <- get_results(
results_thisSeason = example_thisSeason,
seasons = 1L
)
dim(results_combined)
#> [1] 760 9Note that we want to look back across this season (so far) and its predecessor.
game_latest <- calc_game_latest(results = results_combined)
results_filtered <- get_results_filtered(
results = results_combined,
index_game_latest = game_latest,
lookback_rounds = 76L
)
dplyr::glimpse(results_filtered)
#> Rows: 760
#> Columns: 8
#> $ matchday <date> 2024-08-16, 2024-08-17, 2024-08-17, 2024-08-17, 2024-08-17, …
#> $ homeTeam <chr> "MUN", "IPS", "ARS", "EVE", "NEW", "NOT", "WHU", "BRE", "CHE"…
#> $ awayTeam <chr> "FUL", "LIV", "WOL", "BRI", "SOU", "BOU", "AST", "CPA", "MCI"…
#> $ FTHG <dbl> 1, 0, 2, 0, 1, 1, 1, 2, 0, 1, 2, 0, 2, 4, 0, 4, 0, 1, 2, 2, 1…
#> $ FTAG <dbl> 0, 2, 0, 3, 0, 1, 2, 1, 2, 1, 1, 2, 1, 1, 1, 0, 2, 1, 6, 0, 1…
#> $ FTR <chr> "H", "A", "H", "A", "H", "D", "A", "H", "A", "D", "H", "A", "…
#> $ played <lgl> TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, T…
#> $ match <chr> "001", "002", "003", "004", "005", "006", "007", "008", "009"…For reference, we can see the prevailing table.
data_table_current <- example_thisSeason |>
calc_table_current()
data_table_current |>
print_table_current()| Team | Played | GD | Points |
|---|---|---|---|
| ARS | 8 | 12 | 19 |
| MCI | 8 | 11 | 16 |
| LIV | 8 | 3 | 15 |
| BOU | 8 | 3 | 15 |
| TOT | 8 | 7 | 14 |
| CHE | 8 | 7 | 14 |
| SUN | 8 | 3 | 14 |
| CPA | 8 | 4 | 13 |
| MUN | 8 | -1 | 13 |
| BRI | 8 | 1 | 12 |
| AST | 8 | 0 | 12 |
| EVE | 8 | 0 | 11 |
| BRE | 8 | -1 | 10 |
| NEW | 8 | 0 | 9 |
| FUL | 8 | -4 | 8 |
| LEE | 8 | -6 | 8 |
| BUR | 8 | -6 | 7 |
| NOT | 8 | -10 | 5 |
| WHU | 8 | -12 | 4 |
| WOL | 8 | -11 | 2 |
We can now model the strengths of the sides at home and away.
data_model <- results_filtered |>
model_prepare_frame() |>
model_run()
data_model
#>
#> Call:
#> gnm::gnm(formula = count ~ -1 + s + draw, eliminate = match,
#> family = stats::quasipoisson, data = modelframe, start = rep(0,
#> 2 * nTeams + 1))
#>
#> Coefficients of interest:
#> sARS_home sAST_home sBOU_home sBRE_home sBRI_home sBUR_home sCHE_home
#> 3.5887 3.3910 2.6672 2.7102 2.8616 3.0648 3.6256
#> sCPA_home sEVE_home sFUL_home sLEE_home sLIV_home sMCI_home sMUN_home
#> 2.1443 2.2261 2.2417 1.9925 4.2522 3.6907 2.2057
#> sNEW_home sNOT_home sSUN_home sTOT_home sWHU_home sWOL_home sARS_away
#> 3.3055 2.3953 4.3863 1.3325 1.0388 1.2459 3.5010
#> sAST_away sBOU_away sBRE_away sBRI_away sBUR_away sCHE_away sCPA_away
#> 2.2927 2.4675 2.0120 2.4678 -27.1656 2.5494 2.7080
#> sEVE_away sFUL_away sLEE_away sLIV_away sMCI_away sMUN_away sNEW_away
#> 1.9873 2.2584 1.1555 3.4702 2.7270 1.5472 2.3185
#> sNOT_away sSUN_away sTOT_away sWHU_away sWOL_away draw
#> 2.7723 1.7477 1.9700 2.1102 1.5372 0.5316
#>
#> Deviance: 849.1585
#> Pearson chi-squared: 904.8874
#> Residual df: 879Next, we use these team strengths to model future games across the season.
data_parameters_unplayed <- data_model |>
model_extract_parameters()
data_model_parameters_unplayed <- model_parameters_unplayed(
results = results_filtered,
model_parameters = data_parameters_unplayed
)
data_points_expected_remaining <- data_model_parameters_unplayed |>
calc_points_expected_remaining()
calc_points_expected_total(
table_current = data_table_current,
points_expected = data_points_expected_remaining
) |>
knitr::kable()| midName | Exp_Points_Ave |
|---|---|
| Liverpool | 81.65226 |
| Arsenal | 81.16959 |
| Man City | 69.97195 |
| Chelsea | 66.99225 |
| Sunderland | 65.06298 |
| Aston Villa | 60.51185 |
| Newcastle | 59.42471 |
| Bournemouth | 58.81905 |
| Brighton | 58.03840 |
| Crystal Palace | 55.85868 |
| Brentford | 50.92066 |
| Notts Forest | 50.42615 |
| Everton | 48.28917 |
| Fulham | 47.81465 |
| Man Utd | 46.34688 |
| Tottenham | 40.73357 |
| Leeds Utd | 36.83454 |
| Burnley | 34.98184 |
| West Ham | 33.80618 |
| Wolves | 27.85355 |
On this basis, we can see who looks like favourites to win the season.
In order to project the likelihood of these favourites becoming champions, though, we need to simulate many possible outcomes.
number_simulations <- 100000
data_simulate_games <- simulate_games(
data_model_parameters_unplayed = data_model_parameters_unplayed,
value_number_sims = number_simulations,
value_seed = 2602L
)
data_simulate_standings <- simulate_standings(
data_game_simulations = data_simulate_games,
data_table_latest = data_table_current
)
simulate_outcomes(
data_standings_simulations = data_simulate_standings,
value_number_sims = number_simulations
) |>
knitr::kable()| midName | champion | top_four | top_five | top_six | top_half | relegation |
|---|---|---|---|---|---|---|
| Arsenal | 46.9% | 97.8% | 99.1% | 99.6% | >99.9% | <0.1% |
| Liverpool | 46.5% | 98.3% | 99.3% | 99.7% | >99.9% | <0.1% |
| Man City | 4.1% | 68.0% | 80.4% | 88.2% | 99.0% | <0.1% |
| Chelsea | 1.5% | 49.4% | 66.4% | 78.5% | 97.7% | <0.1% |
| Sunderland | 0.5% | 33.8% | 53.3% | 69.1% | 97.0% | <0.1% |
| Aston Villa | 0.2% | 14.5% | 25.9% | 39.1% | 84.7% | <0.1% |
| Bournemouth | 0.2% | 11.2% | 20.6% | 32.2% | 78.5% | <0.1% |
| Brighton | <0.1% | 8.2% | 16.3% | 26.9% | 75.1% | <0.1% |
| Newcastle | <0.1% | 10.8% | 20.4% | 32.2% | 80.5% | <0.1% |
| Crystal Palace | <0.1% | 5.5% | 11.2% | 19.2% | 64.7% | 0.3% |
| Brentford | <0.1% | 0.9% | 2.3% | 4.9% | 33.6% | 1.4% |
| Notts Forest | <0.1% | 0.8% | 2.2% | 4.5% | 30.2% | 2.3% |
| Everton | <0.1% | 0.4% | 1.1% | 2.5% | 21.8% | 3.1% |
| Fulham | <0.1% | 0.3% | 0.9% | 2.0% | 18.8% | 4.1% |
| Man Utd | <0.1% | 0.2% | 0.5% | 1.2% | 14.1% | 6.0% |
| Tottenham | <0.1% | <0.1% | <0.1% | 0.2% | 3.6% | 20.7% |
| Leeds Utd | <0.1% | <0.1% | <0.1% | <0.1% | 0.6% | 44.5% |
| West Ham | <0.1% | <0.1% | <0.1% | <0.1% | 0.2% | 67.3% |
| Burnley | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% | 58.9% |
| Wolves | <0.1% | <0.1% | <0.1% | <0.1% | <0.1% | 91.2% |
Alternatively, this table can be generated, without calculating all
intermediate steps, by running run_simulations.