AFL Predictions

A weekly AFL match prediction system using player statistics, Glicko ratings, and a Keras deep learning model.

How It Works

1. Data: Player stats from Footywire, match results from AFLTables, fixtures from Squiggle — all via the fitzRoy package.
2. Ratings: Glicko2 ratings calculated per team using the PlayerRatings package.
3. Features: ~20 lagged and differential metrics (margins, scoring chains, disposals, ratings, etc.)
4. Model: Dense neural network (256→128→64→32→2) trained on binary win/loss classification. ~71% test accuracy.
5. Output: Win probabilities and margin estimates for each upcoming match.

Setup

Requirements

Install R packages:

install.packages(c(
  "tidyverse", "fitzRoy", "data.table", "PlayerRatings",
  "keras", "plotly", "lubridate", "reshape2", "ggpmisc",
  "magrittr", "jsonlite", "ggpubr", "reactable", "htmltools",
  "sparkline", "formattable", "here"
))
keras::install_keras()

API Key

Betting odds require an API key from the-odds-api.com. Add it to ~/.Renviron:

ODDS_API_KEY=your_key_here

Then restart R or run readRenviron("~/.Renviron").

Running the Pipeline

Set round.no and YEAR in run_pipeline.R, then source it:

source("run_pipeline.R")

This runs the four steps in order:

1. source_code/betting_odds.R — fetch current round odds
2. source_code/AFL_data.R — fetch stats, build features, compute Glicko ratings
3. source_code/prediction_model.R — generate win probabilities and margin estimates
4. source_code/Store_predictions.R — archive to CSV and render output table

After the round is complete, archive results:

source("source_code/Store_betting_odds.R")

Project Structure

run_pipeline.R          # Entry point — run this each round
source_code/            # Main pipeline scripts
  AFL_data.R            # Data fetching, feature engineering, Glicko ratings
  betting_odds.R        # Fetch current round odds from the-odds-api.com
  prediction_model.R    # Load model and generate predictions
  Store_predictions.R   # Archive predictions and render output table
  Store_betting_odds.R  # Archive completed round results (run after round)
functions/              # Shared utilities (auto-sourced at startup)
  config.R              # Project-wide constants (START_SEASON, AFL_TEAMS, etc.)
  team_names.R          # Team name normalisation utilities
  model_data_function.R # Normalise and split data for Keras
  model_training_function.R  # Keras model architecture and training
  line_odds_function.R  # Parse spreads JSON from odds API
  h2h_odds_function.R   # Parse h2h JSON from odds API
  wrangle_fixture.R     # Reshape fixture into home/away format
  simulation_function.R # Monte Carlo margin simulation
  reactable_function.R  # Interactive ratings table
  custom_theme.R        # ggplot2 theme
csv_files/              # Data storage
  AFLstats.csv          # Historical player-level stats (2010–present)
  betting_odds.csv      # Historical match results with betting odds
  round_predictions.csv # All historical predictions
  Team_Ratings.csv      # Latest Glicko ratings snapshot
model/
  model_betless/        # Trained Keras model (betting odds excluded from inputs)
analysis_code/          # Exploratory scripts (not part of the weekly pipeline)
images/                 # Team logo PNGs

Key Constants (functions/config.R)

Constant	Value	Description
START_SEASON	2010	First season included in training data
TRAIN_TEST_SEED	321	Random seed for train/test split
MODEL_PATH	"model/model_betless"	Path to the trained Keras model
AFL_TEAMS	18 teams	Canonical team name list for numeric encoding