The ELO rating system was developed in the 1960s by Hungarian-American physicist Arpad Elo to rank chess players. The beauty of ELO lies in its simplicity: every player (or team) has a rating, and after each match, points are exchanged between competitors based on the outcome and the expected result.
How it works:
ELO is particularly well suited for AFL predictions because:
This tutorial builds an ELO system to predict AFL home team wins, incorporating home ground advantage and tuning parameters for optimal performance.
Note on Draws: Draws are rare in AFL and for this example are treated as home team losses for simplicity.
I used R to implement the ELO model. You can follow along by running the code blocks.
# Import Library
library(fitzRoy)
library(tidyverse)
library(elo)
library(lubridate)
library(tidymodels)
library(xgboost)
library(slider)
library(caret)
library(yardstick)
library(patchwork)
library(vip)
library(knitr)
library(formattable)
library(htmltools)
# Clear Memory
rm(list = ls())Lets use data for all the seasons which the current competing AFL sides have been in the league together. This covers 2012 when GWS joined to 2025. Also as Elo is just win/loss based we only need basic match result data and wont need to make adjustments for the covid interrupted seasons.
We fetch AFL results from 2012-2025 using the fitzRoy package:
# Collect Data
results_df <- fetch_results_afltables(season = 2012:2025)Let’s examine the structure and contents of our data:
print(summary(results_df)) Game Date Round Home.Team
Min. :13960 Min. :2012-03-24 Length:2879 Length:2879
1st Qu.:14680 1st Qu.:2015-06-20 Class :character Class :character
Median :15399 Median :2018-09-06 Mode :character Mode :character
Mean :15399 Mean :2018-12-26
3rd Qu.:16118 3rd Qu.:2022-07-05
Max. :16838 Max. :2025-09-27
Home.Goals Home.Behinds Home.Points Away.Team
Min. : 2.00 Min. : 1.00 Min. : 16.00 Length:2879
1st Qu.:10.00 1st Qu.: 9.00 1st Qu.: 68.00 Class :character
Median :12.00 Median :11.00 Median : 85.00 Mode :character
Mean :12.67 Mean :11.33 Mean : 87.33
3rd Qu.:16.00 3rd Qu.:14.00 3rd Qu.:105.00
Max. :31.00 Max. :28.00 Max. :205.00
Away.Goals Away.Behinds Away.Points Venue
Min. : 1.00 Min. : 1.00 Min. : 14.00 Length:2879
1st Qu.: 9.00 1st Qu.: 8.00 1st Qu.: 63.00 Class :character
Median :11.00 Median :10.00 Median : 79.00 Mode :character
Mean :11.77 Mean :10.49 Mean : 81.14
3rd Qu.:14.00 3rd Qu.:13.00 3rd Qu.: 97.00
Max. :29.00 Max. :25.00 Max. :187.00
Margin Season Round.Type Round.Number
Min. :-138.000 Min. :2012 Length:2879 Min. : 1.00
1st Qu.: -21.000 1st Qu.:2015 Class :character 1st Qu.: 6.00
Median : 5.000 Median :2018 Mode :character Median :12.00
Mean : 6.195 Mean :2019 Mean :12.65
3rd Qu.: 33.000 3rd Qu.:2022 3rd Qu.:19.00
Max. : 171.000 Max. :2025 Max. :29.00 First few rows:
print(head(results_df) )# A tibble: 6 × 16
Game Date Round Home.Team Home.Goals Home.Behinds Home.Points Away.Team
<dbl> <date> <chr> <chr> <int> <int> <int> <chr>
1 13960 2012-03-24 R1 GWS 5 7 37 Sydney
2 13961 2012-03-29 R1 Richmond 12 9 81 Carlton
3 13962 2012-03-30 R1 Hawthorn 20 17 137 Collingw…
4 13963 2012-03-31 R1 Melbourne 11 12 78 Brisbane…
5 13964 2012-03-31 R1 Gold Coa… 10 8 68 Adelaide
6 13965 2012-03-31 R1 North Me… 15 12 102 Essendon
# ℹ 8 more variables: Away.Goals <int>, Away.Behinds <int>, Away.Points <int>,
# Venue <chr>, Margin <int>, Season <dbl>, Round.Type <chr>,
# Round.Number <int>FitzRoy data is generally clean, but let’s do some simple checks to make sure. It’s also a good habit to always inspect your data before modeling, even if you know where the data came from, you will often find small issues that need addressing.
Check team name consistency:
# Preprocessing
# Check consistency
unique(results_df$Home.Team) [1] "GWS" "Richmond" "Hawthorn" "Melbourne"
[5] "Gold Coast" "North Melbourne" "Fremantle" "Footscray"
[9] "Port Adelaide" "Brisbane Lions" "Essendon" "Sydney"
[13] "West Coast" "Adelaide" "Collingwood" "St Kilda"
[17] "Geelong" "Carlton" unique(results_df$Away.Team) [1] "Sydney" "Carlton" "Collingwood" "Brisbane Lions"
[5] "Adelaide" "Essendon" "Geelong" "West Coast"
[9] "St Kilda" "Port Adelaide" "Fremantle" "Melbourne"
[13] "Footscray" "Richmond" "GWS" "Gold Coast"
[17] "Hawthorn" "North Melbourne"Check venue names:
# Check Ground Names
unique(results_df$Venue) [1] "Stadium Australia" "M.C.G." "Carrara"
[4] "Docklands" "Subiaco" "Football Park"
[7] "Gabba" "S.C.G." "Bellerive Oval"
[10] "Blacktown" "Kardinia Park" "Manuka Oval"
[13] "York Park" "Marrara Oval" "Sydney Showground"
[16] "Cazaly's Stadium" "Wellington" "Adelaide Oval"
[19] "Traeger Park" "Jiangwan Stadium" "Eureka Stadium"
[22] "Perth Stadium" "Riverway Stadium" "Norwood Oval"
[25] "Summit Sports Park" "Barossa Oval" "Hands Oval" Check round types:
# Round Type
unique(results_df$Round.Type)[1] "Regular" "Finals" Check for missing values:
# Missing Values
colSums(is.na(results_df)) Game Date Round Home.Team Home.Goals Home.Behinds
0 0 0 0 0 0
Home.Points Away.Team Away.Goals Away.Behinds Away.Points Venue
0 0 0 0 0 0
Margin Season Round.Type Round.Number
0 0 0 0 We already decided to ignore draws as a result outcome, but let’s see how often they occur to check out justification holds.
# Draw Proportion
n_draws <- sum(results_df$Home.Points == results_df$Away.Points)
n_games <- nrow(results_df)
cat("Number of draws:", n_draws, "out of", n_games, "games (", round(100*n_draws/n_games, 2), "% )\n")Number of draws: 21 out of 2879 games ( 0.73 % )Elo models contain parameters that must be set before training - finding optimal values for these parameters is called hyperparameter tuning. This ELO model has three parameters that will impact prediction accuracy which we will tune.
The K-factor controls how reactiveratings are after a result, too high and teams fluctuate wildly, too low and the model is slow to recognise form changes.
Home Ground Advantage (HGA) adds rating points to the home team to reflect their inherent advantage from crowd support, familiarity with conditions, and reduced travel.
Carry Over determines how much of a team’s rating persists between seasons. This accounts for list turnover, coaching changes, and regression to the mean.
To find optimal values, we split our data into training (2012-2023) and validation (2024-2025) sets. We test different parameter combinations on recent unseen games to see which configuration predicts best.
Split data into training (2012-2023) and validation (2024-2025) sets: For this analysis, we use 2024-2025 data for validation, though you could experiment with a larger validation set.
# Training data: 2012-2023
M1_train_data <- results_df %>%
filter(Season <= 2023)
# Validation data: 2024 and 2025
M1_val_data <- results_df %>%
filter(Season >= 2024)
# Check split sizes
cat("Training games:", nrow(M1_train_data), "\n")Training games: 2447 cat("Validation games:", nrow(M1_val_data), "\n")Validation games: 432 Create a function to calculate log loss which we will use as our performance metric:
# Helper Functions
# Calculate log loss
calculate_log_loss <- function(actual, predicted, eps = 1e-15) {
predicted <- pmax(pmin(predicted, 1 - eps), eps)
-mean(actual * log(predicted) + (1 - actual) * log(1 - predicted))
}Create the hyperparameter tuning function:
# Hyper Parameter Tuning Function
tune_elo <- function(k_val, HGA, carryOver, train_df, val_df) {
# Combine train and validation
combined_df <- bind_rows(train_df, val_df) %>%
arrange(Game)
# Run ELO model
elo_model <- elo.run(
as.numeric(Home.Points > Away.Points) ~
adjust(Home.Team, HGA) +
Away.Team +
regress(Season, 1500, carryOver) +
group(Game),
k = k_val,
data = combined_df
)
# Extract all predictions
all_predictions <- as.data.frame(elo_model)
# Get indices for validation games
val_indices <- which(combined_df$Game %in% val_df$Game)
# Filter to validation predictions only
val_predictions <- all_predictions[val_indices, ]
# Get actual outcomes for validation games
val_actual <- combined_df %>%
filter(Game %in% val_df$Game) %>%
mutate(home_won = as.numeric(Home.Points > Away.Points)) %>%
pull(home_won)
# Calculate log loss on validation set
log_loss <- calculate_log_loss(
actual = val_actual,
predicted = val_predictions$p.A
)
return(log_loss)
}We will use grid search to find optimal parameter values. This systematically tests every combination of parameter values across predefined ranges. Our grid tests K-factors from 30-70, HGA from 40-80, and carry-over from 0-0.5, evaluating each combination’s performance on the validation set using log loss, where the smallest log loss combination will be the best parameters.
Define the parameter grid to search:
# Define parameter grid
param_grid <- expand_grid(
k_val = seq(30, 70, by = 2),
HGA = seq(40, 80, by = 2),
carryOver = seq(0, 0.5, by = 0.05)
)Run the grid search (this will take a few minutes):
# Run Hyper Parameter Grid Search
Model1_tuning_results <- param_grid %>%
rowwise() %>%
mutate(
log_loss = tune_elo(k_val, HGA, carryOver, M1_train_data, M1_val_data)
) %>%
ungroup() %>%
arrange(log_loss)View the top 10 parameter combinations:
# Evaluate results
# Print top 10 combinations
Model1_tuning_results %>%
slice_head(n = 10) %>%
print()# A tibble: 10 × 4
k_val HGA carryOver log_loss
<dbl> <dbl> <dbl> <dbl>
1 46 60 0.1 0.572
2 46 62 0.1 0.572
3 46 58 0.1 0.572
4 44 60 0.1 0.572
5 44 58 0.1 0.572
6 44 62 0.1 0.572
7 46 64 0.1 0.572
8 46 56 0.1 0.572
9 48 60 0.1 0.572
10 48 62 0.1 0.572Extract and display the best parameters:
# Best parameters
M1_best_params <- Model1_tuning_results %>% dplyr::slice(1)
# Print Best Parameters
cat("K-factor:", M1_best_params$k_val, "\n",
"Home Ground Advantage:", M1_best_params$HGA, "\n",
"Carry Over:", M1_best_params$carryOver, "\n",
"Validation Log Loss:", round(M1_best_params$log_loss, 4), "\n")K-factor: 46
Home Ground Advantage: 60
Carry Over: 0.1
Validation Log Loss: 0.5719 Set the best hyperparameters:
# Set parameters
# Home Ground Advantage
HGA <- 60
# Regression to the mean
carryOver <- 0.1
# K-Value
k_val <- 46 Run the final model on all data:
# Run Elo Model
Model1 <- elo.run(
as.numeric(Home.Points > Away.Points) ~
adjust(Home.Team, HGA) +
Away.Team +
regress(Season, 1500, carryOver) +
group(Game),
k = k_val,
data = results_df
)# Check the output
summary(Model1)
An object of class 'summary.elo.run', containing information on 18 teams and 2879 matches, with 14 regressions.
Mean Square Error: 0.207
AUC: 0.7314
Favored Teams vs. Actual Wins:
Actual
Favored 0 1
TRUE 572 1249
(tie) 0 0
FALSE 688 370Now the 2025 finals have come and gone we can see where the model ranks all the teams based on their performance over the entire data set.
Get the current ratings for all teams:
# Get final ratings
final_ratings <- final.elos(Model1)
final_ratings <- final_ratings %>%
as.data.frame()
final_ratings %>% arrange(desc(.)) .
Brisbane Lions 1767.721
Geelong 1693.723
Hawthorn 1666.938
GWS 1637.469
Collingwood 1631.229
Adelaide 1624.308
Fremantle 1619.078
Footscray 1593.243
Gold Coast 1583.020
Sydney 1576.269
Port Adelaide 1476.264
Carlton 1445.104
St Kilda 1444.623
Melbourne 1358.091
Essendon 1294.422
North Melbourne 1239.015
Richmond 1222.505
West Coast 1126.978Extract all predictions and rating updates for all the data:
# Get all predictions and ratings
Model1_predictions <- as.data.frame(Model1)
print(head(Model1_predictions)) team.A team.B p.A wins.A update.A update.B elo.A
1 GWS Sydney 0.5854987 0 -26.93294 26.93294 1473.067
2 Richmond Carlton 0.5854987 0 -26.93294 26.93294 1473.067
3 Hawthorn Collingwood 0.5854987 1 19.06706 -19.06706 1519.067
4 Melbourne Brisbane Lions 0.5854987 0 -26.93294 26.93294 1473.067
5 Gold Coast Adelaide 0.5854987 0 -26.93294 26.93294 1473.067
6 North Melbourne Essendon 0.5854987 0 -26.93294 26.93294 1473.067
elo.B
1 1526.933
2 1526.933
3 1480.933
4 1526.933
5 1526.933
6 1526.933Combine predictions with original data:
# Combine with original data
Model1_results <- results_df %>%
bind_cols(Model1_predictions)Calculate accuracy on all games:
# Check accuracy on ALL data
Model1_results <- Model1_results %>%
mutate(
home_won = Home.Points > Away.Points,
predicted_home_win = p.A > 0.5,
correct = home_won == predicted_home_win
)
# Overall accuracy
cat("Overall accuracy:", mean(Model1_results$correct), "\n")Overall accuracy: 0.6728031 Break down performance by each season:
# Performance by season
Model1_by_season <- Model1_results %>%
group_by(Season) %>%
summarise(
n_games = n(),
accuracy = mean(correct),
log_loss = calculate_log_loss(as.numeric(home_won), p.A)
) %>%
arrange(Season)
print(Model1_by_season, n = Inf)# A tibble: 14 × 4
Season n_games accuracy log_loss
<dbl> <int> <dbl> <dbl>
1 2012 207 0.676 0.592
2 2013 207 0.696 0.545
3 2014 207 0.729 0.575
4 2015 206 0.684 0.603
5 2016 207 0.657 0.566
6 2017 207 0.623 0.653
7 2018 207 0.696 0.580
8 2019 207 0.628 0.653
9 2020 162 0.679 0.628
10 2021 207 0.643 0.668
11 2022 207 0.681 0.579
12 2023 216 0.671 0.613
13 2024 216 0.662 0.598
14 2025 216 0.694 0.545This ELO model achieves approximately 67% accuracy in predicting AFL home team wins. The model uses:
The model performs consistently across seasons and provides probabilistic predictions that can be used for tipping competitions.
Hope you enjoyed the brief run through of building an ELO model for AFL match predictions! Feel free to reach out with any questions or suggestions for further improvements.