PitWall Intelligence

Predictive Modeling of Formula 1 Constructor Performance and Sponsorship Value Using Machine Learning

A data-driven Brand Value Index (BVI) for Formula 1 constructors, combining machine-learning performance prediction with SHAP-based explainability — quantifying what is currently a USD 1.8 billion sponsorship market priced largely on perception.

---

Problem

Formula 1's sponsorship market exceeds USD 1.8 billion annually, with a global audience above 400 million viewers across broadcast, streaming, and digital channels. Yet sponsorship valuation in the sport still relies on brand-perception surveys, media-impression estimates, and subjective prestige scoring. There is no publicly available analytical framework that converts on-track performance into an interpretable, comparable, sponsor-facing metric.

Research question

Can explainability techniques applied to structured ML models trained on complete historical race data produce an interpretable composite score of F1 sponsorship value that discriminates between constructors?

Brand Value Index (BVI)

Two-dimensional composite, season-normalised per constructor:

Dimension	Weight	Components
Performance	60%	Predicted championship points · podium probability
Consistency	40%	Reliability indicators · qualifying-to-race delta

Min-max normalisation within each season ensures dominant-era seasons do not suppress midfield-team scores in cross-season comparison.

Headline finding — Sprint 1

Average qualifying gap-to-pole correlates with total constructor points at Pearson r ≈ –0.79 across 112 team-seasons (p < 0.001). Qualifying pace anchors the BVI Performance dimension.

Dataset

Single source — Jolpica-F1 API, an actively maintained mirror of the Ergast Developer API for Formula 1. No Kaggle imports, no third-party aggregators, no synthetic data. Every record is fetched live and cached as JSON for reproducibility.

Focal era: V6 hybrid, 2014–2024 — eleven complete seasons under stable technical regulations, enabling like-for-like cross-season comparison.

Table	Rows	Coverage
races	228	Grands Prix, 2014–2024
results	4,626	Race finishing data
qualifying	4,610	Q1 / Q2 / Q3 session times
constructor_standings	112	Constructor-season finals
driver_standings	247	Driver-season finals
constructors · drivers	—	Team and driver metadata

Methodology

flowchart TD
    A[Jolpica API] --> B[ETL<br/>requests + tenacity · JSON cache]
    B --> C[Preprocessing<br/>DNF typing · qual parsing · gap-to-pole · season norm]
    C --> D[EDA<br/>9 analyses]
    D --> E{Models}
    E --> F[Baseline<br/>Linear Regression · Decision Tree]
    E --> G[Advanced<br/>Gradient Boosting · Logistic + Platt]
    F --> H[Evaluation<br/>5-fold CV · RMSE / R² · AUC-ROC]
    G --> H
    H --> I[SHAP attribution]
    I --> J[BVI synthesis<br/>Performance 60% + Consistency 40%]
    J --> K[Streamlit dashboard]

Loading

Tech stack

Python 3.11 · requests · tenacity · pandas · numpy · SQLite · scikit-learn · shap · matplotlib · seaborn · plotly · streamlit

Environment: Google Colab · Jupyter · VS Code

Sprint plan

#	Window	Focus	Status
1	23 Apr – 4 May	ETL pipeline · preprocessing · EDA	Complete
2	5 May – 18 May	Baseline + advanced models	Complete
3	19 May – 1 Jun	BVI synthesis · SHAP attribution	Complete
4	2 Jun – 15 Jun	Streamlit dashboard	In progress
5	16 Jun – 22 Jun	Report polish · viva prep	Planned

Final report due: 22 June 2026 · Defence: 6 July 2026

Sprint 1 — key findings

1. Qualifying speed is the strongest single predictor of season points. Pearson r ≈ –0.79 across 112 team-seasons (p < 0.001), anchoring qualifying pace as a primary Performance input.

2. The era is defined by sustained dominance. Mercedes won eight consecutive Constructors' Championships (2014–2021), followed by Red Bull (2022, 2023) and McLaren (2024) — concentration that motivates within-season normalisation.

3. Constructor rank volatility varies sharply across the era, feeding the Consistency dimension:

   Constructor	σ (rank)	Tier
   Williams	2.83	Most volatile
   McLaren	2.44	Volatile
   Mercedes	1.04	Stable
   Red Bull	0.92	Stable
   Force India	0.84	Most stable

4. Season concentration ranges from 0.503 (2020, most competitive) to 0.619 (2016, most concentrated), with era-wide mean Gini 0.556.

Numerical findings reflect current Sprint 1 notebook outputs and may be revised after end-to-end re-execution.

Sprint 2 — key findings

Gradient Boosting predicts end-of-season constructor points at held-out 2024 R² 0.976 (five-fold CV R² 0.950), narrowly ahead of a strong Linear Regression baseline (held-out R² 0.956). Permutation importance shows average grid position accounts for almost all of the explained variance — qualifying pace is the spine of constructor performance. Full metrics: reports/sprint2_all_models_metrics.csv; details in notebooks/docs/sprint_2_summary.md.

Sprint 3 — key findings

The two-dimensional Brand Value Index is computed for all 112 team-seasons on a 0–100 scale, combining the Sprint 2 points model with a new calibrated podium-probability classifier — Logistic Regression with Platt scaling, held-out 2024 AUC-ROC 0.928 and Brier 0.071. SHAP attribution on both models confirms starting grid position as the dominant feature. Across 2014–2024 the BVI tracks championship points at a mean Spearman ρ of 0.718 — close enough to be credible, loose enough that the Consistency dimension meaningfully re-ranks teams (2018 Williams, tenth on points, rises to third on reliable grid-to-finish conversion). Scores: data/exports/bvi_scores.csv; figures in reports/; details in notebooks/docs/sprint_3_summary.md.

Repository structure

Pitwall-intelligence/
├── notebooks/
│   ├── 01_etl_jolpica.ipynb         # Jolpica → SQLite, JSON caching + retry
│   ├── 02_preprocessing.ipynb       # cleaning, feature engineering, season normalisation
│   ├── 03_eda.ipynb                 # nine analyses driving Sprint 1 findings
│   ├── 04_features.ipynb            # team_season_features matrix
│   ├── 05_baselines.ipynb           # Linear Regression + Decision Tree
│   ├── 06_advanced.ipynb            # Gradient Boosting · 5-fold CV · permutation importance
│   ├── 07_bvi_shap.ipynb            # podium classifier · SHAP · Brand Value Index
│   └── docs/                        # per-sprint summary write-ups
├── reports/                         # EDA + Sprint 2/3 figures and metrics (PNG, CSV)
├── models/
│   └── gbr_total_points_v1.joblib   # trained Gradient Boosting points model
├── data/
│   └── exports/                     # analytical table snapshots, incl. bvi_scores.csv
├── pitwall.db                       # SQLite analytical store
├── requirements.txt
├── LICENSE
├── .gitignore
└── README.md

Reproducing

git clone https://github.com/DevDharmik/Pitwall-intelligence.git
cd Pitwall-intelligence
pip install -r requirements.txt

Run the notebooks in order in Colab or Jupyter:

1. 01_etl_jolpica.ipynb — populates pitwall.db from the Jolpica API on first run; cached JSON is reused subsequently.
2. 02_preprocessing.ipynb — builds the analytical tables.
3. 03_eda.ipynb — generates the Sprint 1 figures in reports/.
4. 04_features.ipynb — engineers the team_season_features matrix.
5. 05_baselines.ipynb · 06_advanced.ipynb — Sprint 2 models and evaluation.
6. 07_bvi_shap.ipynb — Sprint 3 podium classifier, SHAP attribution, and Brand Value Index.

The repository ships a populated pitwall.db, so notebooks 04–07 can be run directly without re-ingesting from the API.

Author

Dharmik Champaneri — Student ID 20327984 M.Sc. Data Science · University of Europe for Applied Sciences (Berlin / Potsdam) Supervisor: Dr. Humera Noor Minhas Module: Capstone Project · 2026

License

Code, reports, and figures released under the MIT License.