F1 Telemetry Driver Performance Analysis

A machine learning pipeline that models F1 qualifying lap time deltas between driver pairs using telemetry data. The project uses telemetry-derived features from the same qualifying session to estimate lap time differences and identify which driver was faster.

---

Overview

Given two drivers from the same qualifying session, the model answers: based on how each driver drove their fastest lap, who was faster and by how much?

The pipeline:

1. Ingests raw F1 telemetry via the FastF1 API
2. Cleans and aligns multi-driver telemetry to a common distance grid
3. Engineers speed, throttle, and braking features for each lap
4. Builds a labeled dataset of pairwise driver comparisons across multiple race weekends
5. Trains a Random Forest regression model to estimate lap time delta between driver pairs
6. Evaluates performance using grouped train/test splits by race weekend

---

Overview of the Approach

For each qualifying session, the pipeline finds drivers with valid clean laps. For every driver pairing, it:

• extracts the fastest clean lap for each driver
• retrieves and cleans the telemetry data
• aligns both drivers' telemetry to a shared distance grid
• computes per-lap features and pairwise feature differences
• records the actual lap time delta as the target

This project is best interpreted as a same-session telemetry analysis and performance modeling tool, rather than a forward-looking predictor across different sessions.

---

Project Structure

f1/
├── src/
│   ├── data_loader.py       # FastF1 lap retrieval and telemetry extraction
│   ├── preprocess.py        # Telemetry cleaning and grid alignment
│   ├── features.py          # Per-lap and pairwise feature engineering
│   ├── pipeline.py          # Builds one comparison row from a qualifying session
│   ├── visualize.py         # Speed, throttle, brake, and delta plots
│   ├── baseline_model.py    # Optional baseline utilities
│   └── checking_data.py     # Model training and evaluation
├── outputs/
│   ├── tables/              # Generated dataset CSVs
│   └── figures/             # Generated plots
├── run_compare.py           # Main script to build the dataset
├── requirements.txt
└── README.md

---

Features

All pairwise features are computed as Driver A minus Driver B differences.

Feature	Description
mean_speed_diff	Difference in average speed over the lap
top_speed_diff	Difference in maximum speed
min_speed_diff	Difference in minimum speed
std_speed_diff	Difference in speed variability
full_throttle_diff	Difference in percentage of lap at full throttle (>95%)
pct_brake_diff	Difference in percentage of lap spent braking
avg_throttle_diff	Difference in average throttle application
brake_event_diff	Difference in number of distinct braking zones

---

Target

The regression target is the lap time delta between the two drivers:

target_delta_s = lap_time_A_s - lap_time_B_s

• Negative values mean Driver A was faster
• Positive values mean Driver B was faster

The model also supports a derived winner prediction task by checking the sign of the predicted delta.

---

Model and Evaluation

The current model is a RandomForestRegressor trained on telemetry-based pairwise features.

To avoid leaking information between rows from the same race weekend, evaluation uses GroupShuffleSplit, grouping by qualifying session / event. This ensures that entire sessions are held out together during testing.

Reported metrics include:

• MAE — mean absolute error in seconds
• RMSE — root mean squared error
• R² — variance explained
• Winner prediction accuracy — percentage of driver pairs where the model correctly predicts who was faster

Example grouped split output from the current version:

• MAE: 0.2228 s
• RMSE: 0.3222 s
• R²: 0.7229
• Winner prediction accuracy: 98.95%

---

Setup

Requirements

• Python 3.9+
• See requirements.txt for the full dependency list

Installation

git clone https://github.com/Esun4/f1.git
cd f1
python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Cache Setup

FastF1 caches session data locally to avoid repeated API calls. Create a cache directory and update the path in run_compare.py:

fastf1.Cache.enable_cache("path/to/your/cache")

---

Usage

Build the Dataset

python run_compare.py

This loops through the specified qualifying rounds, generates all valid driver pairings, and writes the comparison dataset to outputs/tables/dataset.csv.

Train and Evaluate the Model

python src/checking_data.py

This prints grouped train/test split information, regression metrics, and winner prediction accuracy on the held-out test sessions.

---

Current Results

On held-out qualifying sessions, the current version produces strong same-session performance:

• MAE: 0.2228 s
• RMSE: 0.3222 s
• R²: 0.7229
• Winner prediction accuracy: 98.95%

These results show that telemetry-derived lap features capture a large amount of the variation in within-session lap time differences.

---

Limitations

This project currently uses telemetry features and lap-time targets derived from the same qualifying session. That makes it better suited for same-session performance modeling and telemetry analysis than for true future-session forecasting.

A future extension would be to restructure the pipeline so that earlier-session telemetry is used to predict later-session performance.

---

Dependencies

Key packages:

• fastf1 — F1 session and telemetry data
• pandas, numpy — data manipulation and numerical operations
• scikit-learn — model training and evaluation
• matplotlib — telemetry and delta visualizations

---

License