Premier League Title Prediction

An LSTM-based model that estimates each club's probability of winning the Premier League title, updated week-by-week throughout the season. Outputs are formatted in Kalshi-style percentages — one probability per team per gameweek.

---

How It Works

Data

Historical match results are downloaded from football-data.co.uk — a free source covering every Premier League season from 1993-94 onwards (~32 seasons). No API key required.

From the raw match results the pipeline computes cumulative per-game standings for each team: points, goals, form, and position after every game played.

Features

Each row represents a team's state after their Nth game of the season. Features are split into three groups:

Absolute stats — the team's raw accumulated numbers:

Feature	Description
position	Current league position
form	Numerical score of last 5 results (W=3, D=1, L=0)
won / draw / lost	Cumulative game counts
points	Cumulative points
goalsFor / goalsAgainst / goalDifference	Cumulative goal stats
games_remaining	38 − gameweek

Relative stats — how the team stands compared to the rest of the table at that gameweek:

Feature	Description
points_per_game	points / gameweek — scoring pace, comparable across all stages
points_gap_from_leader	Team's points minus the leader's points (0 for leader, negative otherwise)
max_obtainable_points	points + games_remaining × 3 — mathematical title ceiling

These give the model context within each snapshot, so a team 8 points clear looks very different from a team 8 points adrift, even if their absolute stats are similar.

ELO — a dynamic prior-season strength signal:

Feature	Description
elo	Season-opening quality score, updated each gameweek by live position

ELO Feature

Encodes how strong a team was entering the season, anchored to their previous year's performance:

base_elo = 1 / (prev_position / (1 + trophies_won / possible_trophies))
elo      = base_elo × (100 / current_position)

• prev_position: final league position from the prior season (20 for promoted teams)
• trophies_won: PL + FA Cup + League Cup + European trophies won that season
• possible_trophies: 3 domestic + 1 if in European competition (top-7 finish → Europe)
• current_position: live table position at that gameweek — makes ELO dynamic throughout the season

Higher ELO = stronger team. A title-winning, trophy-laden side that is currently top of the table gets the highest possible score.

Model

A two-layer LSTM with masking and dropout:

Masking → LSTM(64) → Dropout(0.3) → LSTM(32) → Dropout(0.3) → Dense(1)

The output layer produces a raw logit (no sigmoid). At prediction time, logits from all 20 teams are collected and passed through softmax together, treating the problem as a 20-way ranking rather than 20 independent binary predictions. This ensures the model inherently knows only one team can win.

Loss during training: BinaryCrossentropy(from_logits=True) — mathematically identical to sigmoid + BCE but numerically more stable.

Partial-Sequence Training

The model is trained on sequences cut off at gameweeks 5, 10, 15, 20, 25, 30, 35, and 38 — with zero-padding for the remaining weeks. This gives the model 8 training examples per team per season instead of 1, teaching it to make confident predictions from early in the season rather than only at the end.

---

Setup

pip install tensorflow scikit-learn pandas numpy requests joblib

---

Running the Pipeline

Step 1 — Build the training dataset

Downloads all seasons 1993-94 → 2024-25, computes standings, adds all features and labels.

python build_dataset.py

Outputs:

• historical_standings.csv — raw per-game standings
• out.csv — processed training data ready for the model

Step 2 — Train the model

python model_train.py

Outputs:

• best_model.h5 — best LSTM checkpoint (monitored on validation loss)
• scaler.joblib — fitted StandardScaler

Step 3 — Generate 2025-26 predictions

python season_predict.py

Prints Kalshi-style win probabilities for every completed gameweek of the current season, e.g.:

=== Gameweek 15 ===
  Arsenal                41.3%
  Liverpool              22.1%
  Man City               14.8%
  Chelsea                 7.2%
  ...

---

Adding a Newly Finished Season

When a season ends, run add_season.py to append its standings to the training set and patch trophy_data.py with that year's trophy winners. The new features are computed automatically — no other changes needed.

Interactive (will prompt for each trophy winner):

python add_season.py --season 2025

Non-interactive:

python add_season.py --season 2025 \
    --pl Liverpool --fa "Crystal Palace" --lc Liverpool \
    --cl None --el Tottenham --ecl None

Then retrain:

python model_train.py
python season_predict.py

---

Trophy Data

trophy_data.py holds hard-coded historical trophy winners (PL, FA Cup, League Cup, Champions League, Europa League, Conference League) from 1992-93 to 2024-25, keyed by season start year.

All team names must match the football-data.co.uk convention exactly: "Man City", "Man United", "Nott'm Forest", "Sheffield United", etc.

---

File Reference

File	Purpose
build_dataset.py	Downloads all history and builds out.csv
data_aquisition.py	FDCUKDataLoader — free historical CSV fetcher
trophy_data.py	Historical trophy data + ELO formula
prepare_model_data.py	ELO, relative features, form conversion, label assignment
model_train.py	LSTM training with partial-sequence augmentation
season_predict.py	2025-26 Kalshi predictions via softmax across teams
add_season.py	Extend the dataset after a season finishes
constants.py	API key for football-data.org (legacy)