OpenTyreF1: A Data-Driven Mobile Platform for Formula 1 Telemetry Analysis, Tyre-Degradation Modelling and Counterfactual Race-Strategy Simulation

Abstract

Formula 1 publishes one of the richest open telemetry streams in professional sport, yet most public tools built on it are descriptive: they show what happened and stop there. This paper presents OpenTyreF1, a full-stack system that turns that data into decision support on a mobile phone. The system integrates three public sources (Jolpica-F1/Ergast, OpenF1 and FastF1) into one normalised 15-table relational corpus of 237,308 records covering 492 races from 2002 to 2026. On top of this corpus it builds a shared clean-lap filter, a gradient-boosted (XGBoost) lap-time model, a counterfactual stint-strategy simulator, a 1,500-run Monte Carlo outcome estimator, seven pit-wall decision modules and a grounded question-answering assistant, all served by a FastAPI backend to an Expo/React Native app. Evaluated on 12 completely unseen races, the model achieves a mean absolute error of 0.893 s, RMSE of 1.205 s and R2 of 0.978, beating the strongest naive baseline by 0.270 s. An ablation shows that predicting a lap’s offset from its race median, rather than the absolute lap time, cuts error four-fold, a far larger effect than adding features. The results show that, for motorsport data of this size, careful lap admission, target design and leakage-free evaluation matter more than the choice of learner.

Introduction

The paper presents OpenTyreF1, a data-driven Formula 1 race strategy analysis platform that uses machine learning, historical race data, and simulation to help users evaluate alternative tyre strategies and predict potential race outcomes. Although modern Formula 1 cars generate large amounts of telemetry and timing data, most publicly available tools focus on visualising past events rather than answering strategic questions such as how a driver's finishing position might change if they made a pit stop later or selected a different tyre compound.

OpenTyreF1 addresses this limitation by integrating data from public sources, including Ergast, Jolpica-F1, OpenF1, and FastF1, into a unified relational database. The system uses a shared clean-lap filtering process to remove unrepresentative laps and maintain consistency across model training, analysis, and application outputs. An XGBoost machine learning model predicts lap times using historical race data and is evaluated through a whole-race holdout approach to reduce data leakage and provide a more reliable assessment of performance.

The system's central feature is a counterfactual race strategy simulator. Users select a race, driver, tyre compounds, stint lengths, and pit-stop plan. The simulator predicts the driver's lap times, accounts for circuit-specific pit-stop losses, adjusts for safety-car conditions using observed field pace, and compares the hypothetical cumulative race time against competitors' actual performance. Monte Carlo simulation estimates uncertainty in the predicted finishing position, helping users understand the range of possible outcomes rather than relying on a single prediction.

OpenTyreF1 follows a modular client–server architecture with three main layers:

• Frontend: An Expo and React Native mobile application providing live race information, driver statistics, strategy planning, telemetry visualisation, and an AI-assisted interface.

• Backend: A FastAPI service hosting data processing, lap-time prediction, strategy simulation, pit-wall decision modules, and the answer engine.

• Data layer: A relational database using SQLite or PostgreSQL, supported by external Formula 1 APIs and caching mechanisms for efficient data retrieval.

The literature survey compares existing race simulators, machine learning techniques, public data sources, and data-validation methods. It highlights the limitations of conventional visualisation tools, the dependence of some academic simulators on manually specified parameters, and the restricted accessibility of proprietary team software. OpenTyreF1 aims to bridge these gaps by combining reconciled public data, machine learning, counterfactual simulation, uncertainty estimation, and mobile accessibility in one platform.

The system also includes pit-wall decision-support modules and a natural-language assistant designed to provide explanations grounded in actual race data rather than generated text alone. Its data ingestion process is designed to be reproducible, safe under API rate limits, and consistent across multiple sources.

Conclusion

This paper presented OpenTyreF1, an end-to-end system that integrates three heterogeneous public Formula 1 data sources into a single 237,308-record corpus and builds on it a lap-time model, a counterfactual strategy simulator, a Monte Carlo outcome estimator, seven pit-wall decision modules and a grounded assistant, delivered through a cross-platform mobile app. Our central finding is methodological. On data of this size and shape, the decisive choices are not about the learner but about which laps are admitted, how the target is parameterised and how evaluation is partitioned. Reparameterising the target improved MAE four-fold, while nearly doubling the feature count moved it by only 0.033 s. Under a strict whole-race holdout, the model reaches 0.893 s MAE and beats the strongest naive baseline by 0.270 s. Beyond motorsport, the same discipline applies to any product that turns open, noisy event data into recommendations: one shared data definition, a grouped holdout, a matched baseline, declared fallbacks and answers that cite their evidence. OpenTyreF1 shows that these properties can be delivered on free-tier infrastructure, with a 204 MB container and a database small enough to ship inside the repository, so the whole system can be reproduced and audited by anyone with access to the public data. Future work includes correcting the pit-lap rule to work per (driver, lap) and retraining, which would recover about 38% more training data; using Shapley-value attribution [14] to separate tyre-age effects from the fuel-mass proxy; and completing the 2021–2022 backfill so the model can be evaluated across a regulation change. We also plan to estimate the race median online so that the accurate relative model can support live races, and to add a multi-agent field simulation and an energy-deployment (ERS) estimate. In summary, the main contributions of this work are: • A reproducible, idempotent ingestion pipeline that reconciles Jolpica-F1, OpenF1 and FastF1 into one normalised 15-table corpus that can run on either SQLite or PostgreSQL. • A single shared clean-lap filter, imported by the trainer, the analytics services and the API, so that every published pace figure and every model input follows the same definition of a representative lap. • A leakage-free evaluation protocol based on whole-race splits and a matched baseline, which gives an honest estimate of how the model performs on races it has never seen. • A semi-parametric counterfactual simulator that anchors every hypothetical strategy to a real race and to the real lap times of every rival. • Uncertainty-aware decision support through Monte Carlo outcome estimation and seven pit-wall modules that always report the sample size behind each recommendation. • A grounded assistant and a cross-platform mobile client that make this analysis available to fans, students and analysts without proprietary team data. Taken together, these results suggest that useful race-strategy analysis no longer depends on access to private team telemetry. With careful data handling and transparent evaluation, openly published Formula 1 data is sufficient to answer counterfactual strategy questions with measurable accuracy and a clearly stated level of confidence. The same architecture also opens several practical applications. For students and educators, OpenTyreF1 provides a complete, reproducible case study in applied machine learning, in which data ingestion, feature design, leakage-free evaluation and deployment can all be inspected on real public data. For motorsport journalists and commentators, the simulator offers a quick way to test strategy claims made during a broadcast against what the data actually supports. For fans, the mobile client turns post-race discussion from opinion into evidence, since every answer is tied to the laps, stints and pit stops that produced it. For smaller racing teams in junior categories, which rarely have dedicated strategy staff, the same workflow of a shared clean-lap definition, a matched baseline and a declared uncertainty can be adopted with modest engineering effort. Several extensions are planned beyond those listed above. The first is live prediction: the relative target currently needs the race median, so an online estimate built from the opening clean laps of a race and refined as the race progresses would allow the more accurate relative model to be used during live sessions. The second is a multi-agent simulation in which rivals respond to a changed strategy, for example by covering an early pit stop, so that counterfactual positions account for the reactions of the field rather than holding every rival fixed. The third is broader coverage: once the backfilled seasons are fully usable, the model can be retrained across regulation changes and evaluated on how well it transfers from one technical era to the next. Finally, a structured user study with fans and student analysts is planned to measure how well the explanations, sample sizes and uncertainty ranges shown in the app are understood, and whether they improve the quality of strategy judgements made with the tool. Building the system also produced a number of practical lessons that are likely to carry over to similar projects. The most important is that data definitions should be written once and imported everywhere: when the trainer, the analytics services and the API each kept their own idea of a clean lap, their numbers drifted apart, and a single shared filter removed that whole class of inconsistency. The second is that a baseline is only meaningful when it is matched to the target, because a model that predicts offsets from the race median must be compared against the race median itself, not against a weaker global average that it will beat easily. The third is that every fallback should be declared in the output rather than hidden, so that a user can always tell whether a number came from the trained model, from an empirical fit or from a default value. The fourth is that rate limits and data embargoes are design constraints, not afterthoughts, and are best handled by an idempotent ingestion layer that can be stopped and resumed at any point without duplicating or losing records. Finally, OpenTyreF1 is intended to remain open and reproducible. The relational store is small enough to ship with the source code, the serving container runs on free-tier infrastructure, and every figure reported in this paper can be regenerated from the public data sources using the same ingestion, filtering and evaluation code. We hope this makes the system useful not only as a fan and analyst tool but also as a transparent reference implementation for others who wish to build, test and improve data-driven race-strategy methods on openly available Formula 1 data.

References

[1] J. Bekker and W. Lotz, “Planning Formula One race strategies using discrete-event simulation,” Journal of the Operational Research Society, vol. 60, no. 7, pp. 952–961, 2009. [2] A. Heilmeier, M. Graf and M. Lienkamp, “A race simulation for strategy decisions in circuit motorsports,” in Proc. 21st IEEE Int. Conf. Intelligent Transportation Systems (ITSC), 2018, pp. 2986–2993. [3] A. Heilmeier, A. Thomaser, M. Graf and J. Betz, “Virtual strategy engineer: Using artificial neural networks for making race strategy decisions in circuit motorsport,” Applied Sciences, vol. 10, no. 21, art. 7805, 2020. [4] X. Liu and A. Fotouhi, “Formula-E race strategy development using artificial neural networks and Monte Carlo tree search,” Neural Computing and Applications, vol. 32, pp. 15191–15207, 2020. [5] J. H. Friedman, “Greedy function approximation: A gradient boosting machine,” Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001. [6] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2016, [7] pp. 785–794. [8] L. Grinsztajn, E. Oyallon and G. Varoquaux, “Why do tree-based models still outperform deep learning on typical tabular data?” in Advances in Neural Information Processing Systems, vol. 35, 2022. [9] S. Kaufman, S. Rosset and C. Perlich, “Leakage in data mining: Formulation, detection, and avoidance,” in Proc. 17th ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2011, pp. 556–563. [10] D. R. Roberts et al., “Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure,” Ecography, vol. 40, no. 8, pp. 913–929, 2017. [11] C. Newell, “Ergast Developer API: A web service for Formula One motor racing data.” Available: http://ergast.com/mrd/ [12] Jolpica-F1, “An Ergast-compatible API for Formula 1 data.” Available: https://github.com/jolpica/jolpica-f1 [13] OpenF1, “Open-source API for real-time and historical Formula 1 data.” Available: https://openf1.org/ [14] P. Oehrly, “FastF1: A Python package for accessing and analysing Formula 1 results, schedules, timing data and telemetry.” Available: https://github.com/the Oehrly/Fast-F1 [15] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Advances in Neural Information Processing Systems, vol. 30, 2017, [16] pp. 4765–4774. [17] B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap. New York, NY, USA: Chapman & Hall, 1993. [18] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011.

Copyright

Copyright © 2026 Atharva Joshi, Prachi Jain, Ankit Sawalakhe, Kripa Patil, Araju Yelekar, Devshree Khodke. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.