Data Analysis Project: F1 Drivers Dataset

Student Name: BOUDISSA Farouk Radouane
Branch: Systems Security
Group: A2_1
Institution: ENSTA Algiers, Algeria.

Project Overview

This project applies a full data analysis pipeline to the F1 Drivers Dataset, covering 868 Formula 1 drivers and their career statistics from 1950 to the present. Each lab builds on the previous one, progressing from raw data collection and cleaning through statistical analysis, exploratory visualization, and predictive modeling.

Formula 1 choice was mainly subjective as it's my favourite sport, securing a job in its technical teams as a race performance engineer and getting to work with its wide range of real life telemetry is a lifelong dream. This sport was chosen as the dataset domain because its structured scoring system and rich historical records make it ideal for applying every step of the analysis pipeline, from scraping and preprocessing to hypothesis testing and regression.

Dataset

File: formula_one_drivers.csv
Source: Wikipedia — List of Formula One Drivers
Records: 868 drivers
Original features: 11 columns (Driver, Nationality, Seasons, Championships, Race_Entries, Race_Starts, Pole_Positions, Race_Wins, Podiums, Fastest_Laps, Points)
Engineered features (added across labs): Win_Rate, Podium_Rate, Pole_Rate, FastLap_Rate, Points_Per_Entry, Years_Active, Decade, Champion, Active. bringing the working dataset to 22 features

Labs Completed

Lab 1 : Data Collection

Loaded the raw F1 dataset and conducted a first structural inspection using pandas
Identified column types, missing values, and inconsistent formatting in the Seasons column
Noted that Points contains zeros for the majority of the dataset, a pattern that would shape all subsequent analysis
Produced a first look at the dataset dimensions: 868 rows × 11 columns
Flagged that numeric fields were stored as strings in the raw source, requiring type casting in preprocessing

Lab 2 : Inferential Statistics

Explored inferential statistics using the F1 dataset as a storytelling lens to make each concept tangible

Basic & Conditional Probability

Computed the probability of randomly selecting a British driver: ~18.9%, nearly 1 in 5 F1 drivers are British
Given a driver is a World Champion, the probability they are British: ~29.4%

Normal Distribution & Skewness

Plotted the distribution of career Points: skewness of 9.55, one of the most extreme right-skewed distributions possible
More than half of all 868 drivers in the dataset have zero career points

Z-Score & Driver Ranking

Calculated Z-scores for career points across all drivers, producing an all-time statistical ranking
Lewis Hamilton: Z-score ≈ +16.4, so far above average he sits outside any standard distribution scale
Demonstrated that Z-scores are a fairer cross-era comparison tool than raw points

Central Limit Theorem

Simulated 1000 sample iterations across sizes n = [1, 2, 5, 10, 100, 200]
Confirmed CLT: even with extremely skewed Points data, sample means converge to a normal distribution at n ≥ 100

Hypothesis Testing

Z-Test (British drivers): p-value ≈ 0 → British drivers score significantly more than average (mean: 64.4 vs 55.8 overall)
T-Test (Brazilian drivers): p-value < 0.05 → Brazilian drivers average over 6× the career points of a typical driver. Driven by Senna, Piquet, and Fittipaldi

Chi-Square Tests

Goodness of fit: Driver entries are NOT equally distributed across decades (p ≈ 0), the 1950s–60s had far more unique drivers as F1 was open to amateur entrants but it became far more selective as time went on
Independence test: Nationality and champion status are NOT independent (p < 0.05), certain nationalities produce champions at a statistically disproportionate rate (eg. Brazil)

Lab 3 : Web Scraping

Collected supplementary F1 data using the Wikipedia Python library
Scraped driver biography pages to enrich the dataset with additional context unavailable in the base CSV
Parsed and cleaned the scraped content using pandas string operations
Cross-referenced scraped data against the existing dataset to verify consistency
Identified discrepancies in driver names between the Wikipedia article titles and the dataset's Driver column, requiring manual normalization for a subset of entries

Lab 4 : Data Preprocessing

Applied a full preprocessing pipeline to prepare the dataset for analysis and modeling

Cleaning

Cast all numeric columns from string/object to float, replacing non-numeric entries with NaN then filling with 0
Standardized the Seasons column from irregular string ranges (e.g. "1962–1963") into a structured list format
Removed one duplicate entry identified via df.duplicated()

Feature Engineering

Years_Active: extracted from the Seasons ranges, counts the number of distinct seasons each driver competed
Decade: debut decade derived from the first year in Seasons, used as a categorical era indicator
Champion: binary flag (1 if Championships > 0, else 0)
Active: binary flag based on whether the driver's last season is 2023 or 2024
Rate features: Win_Rate, Podium_Rate, Pole_Rate, FastLap_Rate, Points_Per_Entry, all computed as per-entry ratios to enable fair cross-era comparisons

Observations

After cleaning, no missing values remained in any numeric column
The Active flag confirmed exactly 20 active drivers which is consistent with F1's 20-seat grid
Points_Per_Entry proved more informative than raw Points for comparing drivers across different career lengths

Lab 5 : Exploratory Data Analysis

The EDA replicated the methodology of the housing dataset study from the course, adapted to the structure and domain logic of the F1 dataset. Career Points served as the primary target.

Target Distribution

Points is extremely right-skewed: the median is 0, the mean is ~55.8, and the maximum exceeds 4,600
A log transformation was flagged as necessary for any modeling step

Numerical Feature Distributions

Most numerical features (Wins, Podiums, Pole_Positions) follow the same heavy right-skew as Points: a small elite dominates, the majority scores zero
Years_Active is more normally distributed, centered around 3–4 seasons which means that the typical F1 career is short

Correlation Analysis

Raw correlation: Podiums has the strongest correlation to Points across the full dataset
Zero-filtered correlation: when zero-scorers are excluded, rate features (Win_Rate, Podium_Rate) gain dramatically in correlation strength revealing that they are meaningful only for competitive drivers
Two distinct feature clusters identified: volume metrics (Podiums, Race_Wins, Race_Entries) and rate metrics (Win_Rate, Podium_Rate, Points_Per_Entry) each capturing a different dimension of performance

Feature-to-Feature Relationships

Volume metrics intercorrelate heavily: a driver with many wins almost certainly has many podiums and entries, they are not independent pieces of information
Rate metrics intercorrelate too, but less tightly: a high Win_Rate does not always imply a high Fastest_Lap_Rate, since these measure different aspects of speed

Categorical Analysis

Champion flag: champions are statistically distinct on every single metric, both volume and rate
Decade: the 2010 points system change inflates raw Points for modern drivers; cross-era comparison using raw Points is misleading
Nationality: UK and USA dominate driver counts historically; modern representation has shifted toward Europe
Active flag: exactly 20 active drivers, consistent with the grid limit; their Points values are naturally lower as careers are still ongoing

EDA Summary Table

Dimension	Insight
Target (Points)	Extreme right-skew; log-transformation recommended for modelling
Volume vs. Efficiency	Two distinct feature families capture different performance dimensions
Strongest raw predictor	Podiums (r ≈ 0.97 with Points)
Zero-filtering effect	Rate features gain strong correlation when backmarkers are excluded
Multicollinearity	Volume metrics form one tight cluster; rate metrics form another
Era bias	Decade confounds raw Points; Win_Rate or Podium_Rate preferred for cross-era comparison
Best categorical separator	Champion flag, statistically distinct on every metric

Lab 6 : Linear Regression Modeling

Regression was applied to predict career Points (log-transformed) from 7 features selected directly from the EDA findings.

Feature Selection

Included: Podiums, Race_Wins, Pole_Positions, Fastest_Laps, Race_Entries, Years_Active, Championships
Excluded: Rate features (Win_Rate, Podium_Rate) derived from Points itself, would cause data leakage; Race_Starts nearly collinear with Race_Entries

Simple Linear Regression (Podiums only baseline)

Even with a single predictor, the model explains a large share of variance in career Points
Confirms the EDA finding: career points totals are, to a first approximation, a direct function of how many times a driver stood on the podium

Multiple Linear Regression (OLS all 7 features)

All seven coefficients are positive domain-consistent, nothing in F1 penalizes a driver's point tally
Podiums and Race_Entries carry the largest coefficients, reflecting two distinct paths to accumulating points: peak performance and career longevity
Residuals are approximately centered at zero; mild heteroscedasticity appears at the high end where champion-level outliers (Hamilton, Schumacher) sit

Ridge Regression (L2)

Ridge shrinks inflated OLS coefficients caused by the multicollinearity between Podiums, Race_Wins, and Race_Entries identified in the EDA
Slight improvement in generalization over OLS, with all features preserved. appropriate since the EDA confirmed that every selected feature contributes meaningfully to Points

Lasso Regression (L1)

Lasso zeros out some volume features (typically Pole_Positions or Fastest_Laps depending on the fold), confirming that they are redundant given Podiums and Race_Wins
This automatic feature selection is directly consistent with the two-cluster multicollinearity structure observed in Lab 5

Cross-Validation (5-Fold)

All three models maintain consistent R² across all 5 folds with low inter-fold variance
The relationship between these features and Points is stable across different subsets of the 868-driver dataset. the model is not overfitting a particular split

Key Observations

Points in F1 are highly predictable from career statistics. the sport's fixed, rules-based points system leaves very little noise between performance and outcome
Two paths to points: the model quantifies what F1 fans intuitively know, a driver can accumulate points by being exceptionally fast (high Podiums, Race_Wins) or by competing for a very long time (high Race_Entries, Years_Active)
Era bias persists: the 2010 points system change means modern drivers score more raw points than equally dominant drivers from earlier eras; the residual plot shows the model systematically underestimates modern drivers and overestimates older ones
Champions distort the high end: the 34 world champions in the dataset are leverage points that create increasing residual spread at high predicted values. they belong in the model, but their influence on the regression line is disproportionate to their number

Key Findings Across the Project

Total drivers analyzed: 868, spanning 1950 to 2024
Median career points: 0, more than half of all F1 drivers never scored a single point
Career points skewness: 9.55, extreme right-skew consistent across all analysis steps
Lewis Hamilton Z-score: +16.4, the highest in the dataset by a significant margin
World Champions: 34 out of 868 (~3.9%), statistically rare and distinct on every metric
British drivers: ~18.9% of all drivers and ~29.4% of all champions
Brazilian drivers average ~6× the career points of a typical driver
The 1950s–60s account for a disproportionate share of all driver entries, F1 was open to amateurs
Podiums is the single strongest predictor of career Points (r ≈ 0.97)
All regression models achieved R² > 0.90 on both test set and 5-Fold CV, confirming that career point totals are structurally predictable from observable statistics

Files Submitted

analysis.py — Lab 0 Introduction
lab_1.ipynb — Data collection and first inspection
lab_2.ipynb — Inferential statistics
lab_3_4.ipynb — Web scraping with Wikipedia, Data preprocessing and feature engineering
lab_5.ipynb — Exploratory data analysis
f1_app.py — Streamlit data visualisation
lab_7.ipynb — Linear regression and modelling
formula_one_drivers.csv — Dataset
app.py — Streamlit dashboard (full pipeline deployment)
README.md — This file

How to Run

# Install dependencies
pip install pandas numpy matplotlib seaborn scipy statsmodels scikit-learn streamlit wikipedia jupyter

# Run any lab notebook
jupyter notebook lab_6.ipynb

# Launch the Streamlit dashboard
streamlit run app.py

Libraries Used

pandas — Data manipulation, feature engineering, cross-tabulation
NumPy — Numerical computing and random sampling
matplotlib — Data visualization
seaborn — Statistical graphics
SciPy — Statistical tests (chi-square, t-test, z-critical values)
statsmodels — Z-test implementation
scikit-learn — Linear, Ridge, and Lasso regression; train/test split; cross-validation; StandardScaler
Wikipedia — Web scraping for supplementary data collection
Streamlit — Interactive dashboard deployment

Submission Status: Complete ✓

Name		Name	Last commit message	Last commit date
.devcontainer		.devcontainer
Lab 0		Lab 0
Lab 1		Lab 1
Lab 2		Lab 2
Lab 3_4		Lab 3_4
Lab 5		Lab 5
Lab 6		Lab 6
Lab 7		Lab 7
streamlit		streamlit
F1DriversDataset.csv		F1DriversDataset.csv
Guide.md		Guide.md
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation