Fantasy Football Prediction System

This system predicts Fantasy Premier League (FPL) player performance using machine learning models trained on historical data. It also selects an optimal 15-player squad using point predictions, following official FPL rules.

Overview

The pipeline performs:

1. Training position-specific regression models using historical season data.
2. Evaluation using standard metrics (MAE, RMSE, R²).
3. Inference to predict points for a given season and generate an optimal squad.

Features

1. Position-Specific Models

• Separate models for each position (GK, DEF, MID, FWD)
• Model selection per position based on validation performance
• Supported models:
  • Linear Regression
  • Ridge Regression (with cross-validation)
  • Lasso Regression (with cross-validation)
  • ElasticNet Regression (with cross-validation)

2. Feature Engineering

• Player performance metrics:
  • Match statistics (minutes, goals, assists, etc.)
  • Bonus point indicators (BPS, influence, creativity, threat)
  • Form indicators (rolling averages)
• Team and opposition factors:
  • Team strength indicators
  • Home/away performance
  • Opposition difficulty ratings

3. Team Optimization

• Maximizes predicted total points
• Enforces FPL constraints:
  • Budget limit (default: £100m)
  • Squad composition:
    • 2 Goalkeepers
    • 5 Defenders
    • 5 Midfielders
    • 3 Forwards
  • Maximum 3 players per team
• Provides detailed team analysis:
  • Position-wise breakdown
  • Team distribution
  • Value distribution

Requirements

• Python 3.8+
• All dependencies are listed in requirements.txt

requirements.txt

pandas
numpy
scikit-learn
pyarrow
pulp

---

Installation

git clone https://github.com/SanketAdlak/FPLpredictor.git

python -m venv venv
venv\Scripts\activate             # On Windows
# OR
source venv/bin/activate          # On macOS/Linux

pip install -r requirements.txt

Project Structure

src/
├── data/
│   └── 2021-22/
│   └── 2022-23/
│   └── 2023-24/
│   └── 2024-25/
│       ├── players_raw.csv
│       └── merged_gw.csv
├── models/              # Auto-created during training
├── processed/           # Auto-created feature storage
├── train.py             # Train or fine-tune model
├── eval.py              # Evaluate model on test season
├── infer.py             # Predict points and select squad
├── requirements.txt
└── README.md

How to Use

1. Train the Model

python train.py

Saves:

• models/*.joblib (1 per position)
• models/config.json
• processed/features.parquet

Quick Start (Skip Training)

• Pretrained models are included in the repository.
• You can skip train.py and directly run eval.py and infer.py

---

2. Evaluate Model Performance

python eval.py

Evaluation Results

MAE by position:
  GK: 1.894
 DEF: 2.182
 MID: 1.858
 FWD: 2.587
 ALL: 2.046

MSE by position:
  GK: 6.536
 DEF: 7.267
 MID: 8.156
 FWD: 11.766
 ALL: 8.053

R² by position:
  GK: 0.090
 DEF: 0.001
 MID: 0.095
 FWD: 0.082
 ALL: 0.077

---

3. Predict and Optimise Squad

python infer.py data/2024-25/merged_gw.csv --players_raw data/2024-25/players_raw.csv --season 2024-25 --out predictions_2024-25.csv

Optimal Squad Output

name	team	position	cost_m	predicted_points	Starting
Daniel Muñoz	Crystal Palace	DEF	5.2	5.61	True
Diogo Dalot Teixeira	Man Utd	DEF	5.0	4.35	True
Marc Cucurella	Chelsea	DEF	5.4	4.68	True
Nikola Milenković	Nott'm Forest	DEF	5.1	4.78	True
Jørgen Strand Larsen	Wolves	FWD	5.4	4.15	True
Dean Henderson	Crystal Palace	GK	4.6	4.16	True
Bruno Borges Fernandes	Man Utd	MID	8.6	4.77	True
Ismaïla Sarr	Crystal Palace	MID	5.7	4.49	True
Jarrod Bowen	West Ham	MID	7.6	4.40	True
Mohamed Salah	Liverpool	MID	13.8	4.90	True
Morgan Rogers	Aston Villa	MID	5.6	4.57	True
Jurriën Timber	Arsenal	DEF	5.6	4.29	False
Liam Delap	Ipswich	FWD	5.6	3.68	False
Yoane Wissa	Brentford	FWD	6.6	3.82	False
Matz Sels	Nott'm Forest	GK	5.1	3.97	False

---

This above output is the prediction for Gameweek 32.

Note: For visualizations and plots (e.g., feature importance, prediction vs. actual), you can use the provided data in a Jupyter Notebook.
The notebook environment is ideal for exploring results interactively.
Before running notebook, uncommnet the packages from requirements.txt and install all of them.

---

Input Data Format

Each season folder (e.g., data/2024-25/) contains:

1. players_raw.csv

name	position	team	now_cost	element_type	...
Haaland E.	FWD	MCI	125	4	...

2. merged_gw.csv

name	round	minutes	goals_scored	assists	...
Haaland E.	1	90	1	0	...

Model Features

• Rolling stats from previous games (e.g., 5-game window)
• Match stats: goals, assists, minutes, clean sheets
• Advanced stats: BPS, influence, creativity, threat
• Context features: home/away, opponent strength

---

Team Optimization Rules

• Max budget: £100.0m

• Squad size: 15 players

• Max 3 players per club

• Valid formation enforced

  • 2 GKs, 5 DEFs, 5 MIDs, 3 FWDs
  • At least 3 DEF, 2 MID, 1 FWD in starting XI

Limitations

1. Model Assumptions:

   • Past performance indicates future results
   • Linear relationships between features
   • Independent feature contributions

2. Data Limitations:

   • Limited historical data
   • Player transfers between seasons
   • Team strategy changes

3. External Factors:

   • Injuries and suspensions
   • Team formation changes
   • Weather conditions

Future Enhancements

• Fixture-aware prediction
• Transfer suggestion engine
• Player form tracking
• Web-based UI