---

🏎️ Formula 1 Race Outcome Prediction

Predicting whether a Formula 1 driver will finish on the podium (top 3) using historical data, advanced feature engineering, and machine learning.

---

📌 Table of Contents

• Introduction
• Key Terminology
• Problem Statement
• Data Sources
• Methodology
• Exploratory Data Analysis
• Statistical Testing
• Feature Engineering
• Modeling
• Deployment
• How to Run
• Conclusion

---

Introduction

The world of Formula 1 is rich with data—qualifying rounds, race conditions, team strategies, driver performances, and track histories. This project leverages that data to predict one thing: will a driver finish on the podium?

---

🏁 Key Terminology

• Driver — The individual competing in the race. Each F1 team has two drivers.
• Constructor (Team) — The organization that builds and races the car. Examples: Mercedes, Ferrari, Red Bull Racing.
• Grand Prix (Race, Round) — A single event in the F1 calendar, typically held over a weekend, consisting of practice sessions, qualifying, and the main race.
• Qualifying — A session that determines the starting grid for the race. A better qualifying position often improves race performance.
• Grid Position: The starting position of a driver in the race.
• Podium — The top 3 finishers in a race — 1st, 2nd, and 3rd place. These are the drivers who physically stand on the podium after the race and receive trophies.
• Pole Position — The first position on the starting grid, awarded to the fastest qualifier.
• Pit Stop — When a driver enters the pit lane to change tyres or fix minor issues. Time-consuming, but sometimes strategically vital. The pit stop itself ideally takes 2-3 seconds, but the whole process of entering and exiting the pits lasts about 20-25 seconds.
• DNF (Did Not Finish) — When a driver does not complete the race due to a crash, mechanical failure, or other issue.

---

Problem Statement

Goal: Predict whether a driver will finish on the podium using a historical dataset.

This is a binary classification problem with significant class imbalance (only ~15% of drivers finish in the top 3).

---

Data Sources

• Main datasets: https://www.kaggle.com/datasets/rohanrao/formula-1-world-championship-1950-2020
• Pit Stops: https://www.kaggle.com/datasets/akashrane2609/formula-1-pit-stop-dataset
• Weather API and race metadata
• Historical driver/team performance

---

Methodology

1. Data cleaning & merging
2. Feature engineering (both static and temporal)
3. Statistical testing
4. Model training with hyperparameter tuning
5. Evaluation with real-world test set (2024 races)
6. Deployment-ready pipeline + API

---

Exploratory Data Analysis

Exploration included:

• Class imbalance check
• Grid position impact
• Team and driver podium rates
• Circuit-based performance
• Weather condition summaries
• Global distribution of F1 circuits

---

Statistical Testing

We used:

• Chi-Squared Tests: For independence of categorical features.
• Mann-Whitney U Tests: For differences in feature distributions across classes.
• Results informed feature selection.

---

Feature Engineering

Key features:

• Driver Experience (race count)
• Recent Performance (last 3 races)
• Rolling Average Finish
• Constructor Podium Rate
• Track-Specific Averages
• Weather Flags (wet, windy, hot, cold)
• Binary-Encoded Categorical Features

All feature engineering is encapsulated in a reusable F1DataPreprocessor transformer.

---

Modeling

1. Baseline Models

• Logistic Regression
• Random Forest

2. Class Imbalance Handling

• Cost-sensitive learning
• SMOTE and over-sampling

3. Advanced Models

• HistGradientBoostingClassifier
• LightGBM, XGBoost, CatBoost
• Ensembles with Voting and Stacking

4. Hyperparameter Tuning

• Optuna with AUC & F1 scores

---

🏁 Final Model

Best Model: Random Forest with Optuna-tuned hyperparameters Test AUC: 937 Test F1: 0.72 Test Precision: 0.64

---

Deployment

Components

• Custom Transformer (F1DataPreprocessor)

• Custom Pipeline (F1Pipeline)

• Joblib Model Saving:

  joblib.dump(preprocessor, "f1_preprocessor.pkl")
  joblib.dump(model, "models/model.pkl")

FastAPI Inference Server

A FastAPI service is provided for predicting podium chances in real-time.

Dockerized API

The API is dockerized for easy deployment:

docker build -t f1-predictor .
docker run -p 8000:8000 f1-predictor

---

🛠️ How to Run

1. Clone the repo:

   git clone https://github.com/wsiqz/formula-1.git
   cd formula-1

2. Install dependencies:

   pip install -r requirements.txt

3. Run the pipeline notebook:

   notebooks/f1.ipynb
   

4. Train and export the model:

   joblib.dump(preprocessor, "f1_preprocessor.pkl")
   joblib.dump(model, "models/model.pkl")

5. Start FastAPI server:

   uvicorn app.main:app --reload

---

🧾 Conclusion

This project demonstrates how domain knowledge, careful feature engineering, and rigorous modeling techniques can be combined to solve real-world predictive problems—even in complex, dynamic environments like Formula 1.

---