💻 computer science

Baseline-relative evaluation of lap-time and tyre-degradation models for Formula 1 race strategy

This study evaluates Formula 1 lap-time and tyre-degradation models against rigorous baseline forecasts using 167,083 laps, demonstrating that their predictive advantage stems primarily from modeling residuals relative to the previous lap rather than absolute values, and that fair baseline construction is critical for accurately assessing strategy-relevant performance.

Original authors: T. Shakila, K. Baalaji

Published 2026-10-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: T. Shakila, K. Baalaji

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of Formula 1 racing, the difference between winning and losing often comes down to a single decision: when to stop the car and change its tyres. Teams employ strategists who must predict two things with extreme precision. First, they need to know how fast the car will go on the very next lap. Second, they must estimate how much that speed will drop as the rubber wears down over time. These predictions are not guesses; they are calculated using complex computer models trained on years of race data. If these models are accurate, they can guide a team to a victory. If they are flawed, even the most sophisticated race plan will fail. For years, researchers have published studies claiming their models are highly accurate, often citing tiny errors in their predictions. However, a critical piece of the puzzle has been missing from these reports: a comparison against the simplest possible guess. In a sport where a car's speed on one lap is almost identical to the next, simply predicting that the next lap will be the same as the last one is a surprisingly strong baseline. Without knowing how well this simple method performs, it is impossible to tell if a complex computer model is actually adding any value or just repeating what is already obvious.

A new study by researchers at the Bharath Institute of Science and Technology sets out to fix this gap. They gathered a massive dataset of over 167,000 laps from public Formula 1 records spanning several seasons. Their goal was to test advanced machine learning models against a "simple" forecast, which is simply the act of assuming the next lap will take exactly as long as the previous one. The researchers built a sophisticated model using a technique called gradient boosting, which helps computers learn patterns from data by combining many simple decision rules. They trained this model on data from 2019 to 2022 and then tested it on races from 2023, 2024, and 2025 to see if it could predict lap times better than the simple "repeat the last lap" method. The results were clear: the advanced model did indeed outperform the simple guess. On average, the model reduced the prediction error by about 15 percent compared to the simple method. This means that for every lap, the model's guess was closer to the actual time by a small but significant margin.

However, the researchers discovered that the size of this advantage depended entirely on how the comparison was set up. When they compared the model to the simple guess across the entire race, the model looked incredibly strong. But a closer look revealed a flaw in the comparison. The simple guess sometimes failed because it tried to predict a lap on fresh tyres using data from a lap on worn tyres, such as immediately after a pit stop. Since the model knew a new tyre had been fitted, it had an unfair advantage over the simple guess in those specific moments. When the researchers corrected the comparison to ensure the simple guess also knew about the tyre change, the model's advantage shrank from 15 percent to about 12 percent. This 12 percent figure is considered the true measure of the model's skill. The study also found that the model's success was not spread evenly across the race. It was most effective during the first few laps after a pit stop, where it improved predictions by over 30 percent. In the middle of a race, when the car is running steadily, the model's edge was much smaller and sometimes not statistically significant at all.

The study also tackled the question of tyre degradation, or how much speed is lost as the tyre ages. The researchers found that the way they defined "degradation" in their computer model fundamentally changed what the results meant. They used a method where the model predicted the extra time a lap would take compared to a fresh tyre. They discovered that because this calculation is mathematically tied to the lap time itself, any success the model had in predicting degradation was actually just success in predicting the lap time. The model was not learning anything new about how tyres wear out; it was simply getting better at guessing the next lap's speed. This finding suggests that previous studies claiming to predict tyre wear might have been measuring lap-time prediction instead. The researchers also tested whether their model needed to be retrained every time the rules of the sport changed, such as when new car designs were introduced. They found that a model trained on older data performed just as well as one trained on newer data, and combining all the history actually made the model stronger. This implies that teams do not need to discard years of data when regulations change; the patterns of how a car behaves from one lap to the next remain consistent enough to be useful.

Ultimately, this research offers a clearer view of what these prediction models can and cannot do. It confirms that advanced computer models can predict lap times better than a simple guess, but only when the comparison is fair and accounts for the specific conditions of the race, like tyre changes. The models are particularly useful during the chaotic moments of a race, such as right after a pit stop, but they offer less help when the car is running smoothly. The study also serves as a warning for how these models are evaluated. It shows that without a fair baseline and a careful definition of the problem, researchers can easily overstate their success. For the teams that rely on these forecasts, the lesson is practical: use these tools to make decisions during transitions, like pit stops, but do not rely on them to predict long-term tyre wear in a single step. The study does not claim to have solved the mystery of tyre degradation, but it has successfully cleared away the confusion surrounding how we measure the accuracy of these predictions, ensuring that future strategies are built on solid ground rather than statistical illusions.

Technical Summary: Baseline-Relative Evaluation of Lap-Time and Tyre-Degradation Models for Formula 1 Race Strategy

Problem Statement
Formula 1 race strategy relies on forecasts of next-lap time and tyre degradation to make sequential decisions under time constraints. While numerous studies apply machine learning to predict lap times, their reported accuracy is often uninterpretable. These studies typically report Mean Absolute Error (MAE) or coefficients of determination ( $R^2$ ) against other learned models but fail to compare their performance against a simple "naive" forecast (e.g., repeating the previous lap's time) on the exact same held-out data. Because consecutive green-flag laps by the same driver are highly autocorrelated (differing by only tenths of a second), a simple forecast is already a strong baseline. Without this comparison, reported errors cannot be separated from the inherent predictability of the time series, rendering claims of model superiority ambiguous. Furthermore, existing literature often lacks a fair construction of this baseline, ignoring structural factors like pit stops and tyre compound changes that a human strategist would account for.

Methodology
The study utilizes 167,083 laps from the 2019–2025 Formula 1 seasons, retrieved via the public FastF1 interface. The dataset is filtered for green-flag laps with accurate timing, excluding pit-in/out laps and outliers exceeding 110% of a driver's expanding median pace.

Models: The primary model is a Gradient-Boosted Tree (LightGBM) trained on 2019–2022 data and frozen for evaluation on the 2023 hold-out set (and subsequently 2024–2025). A GRU network is also evaluated for degradation.
Target Formulation: The study compares two target formulations:
1. Absolute: Predicting the raw lap time.
2. Residual: Predicting the change from the previous lap (i.e., $y(t) - y(t-1)$ ).
Baselines: Two simple baselines are employed:
1. Persistence: The immediately preceding retained lap time.
2. Stint-Aware Persistence: A fairer baseline that carries a compound-transition offset across pit stops, acknowledging that a strategist knows the new tyre compound immediately after a stop.
Evaluation Protocol: Inference is conducted at the race level, not the lap level, to respect the correlation of laps within a race. Performance is measured by Mean Absolute Error (MAE) and compared using paired race-level t-tests, sign tests, and Diebold–Mariano tests. The study also employs a nested ablation to isolate the effects of feature blocks (conditions, pace, traffic) and target formulations.
Causal Verification: Two automated tests verify that features do not leak future information or depend on the outcome of the predicted lap itself.

Key Contributions

Baseline-Relative Evaluation: The study provides the first systematic comparison of F1 lap-time models against a simple forecast on identical held-out laps, establishing a "simple floor" for accuracy.
Fair Baseline Construction: It demonstrates that simple baselines can be unfairly handicapped. A standard persistence forecast compares laps on different tyres across pit stops, artificially inflating the model's apparent advantage. A stint-aware baseline, which accounts for compound offsets, yields a more defensible measure of skill.
Target Formulation Dominance: The study shows that the choice of target formulation (residual vs. absolute) is more critical than the feature set. Residual targets consistently outperform absolute targets, while absolute targets often fail to beat the simple baseline.
Structural Limitation of Degradation Targets: It reveals a structural property of stint-referenced degradation targets (lap time minus a within-stint constant). At a one-step horizon, the persistence forecast for degradation is mathematically identical to the persistence forecast for lap time. Consequently, any lap-time model's degradation error is identical to its lap-time error, meaning such targets cannot distinguish between a model that has learned about tyre degradation and one that has only learned about lap-time autocorrelation.
Regulation Transferability: The study investigates whether models trained on pre-2022 regulations transfer to the post-2022 ground-effect era, finding that accumulating historical data (including superseded regulations) improves performance compared to training only on the new era.

Results

Lap-Time Prediction: Against a standard persistence baseline, the gradient-boosted model achieves a 15.1% reduction in MAE. However, against the fairer stint-aware baseline, the advantage reduces to 12.1% (statistically significant, $p=0.0065$ ).
Stratified Performance: The model's skill is concentrated in transitions. It achieves a 30.6% improvement in the first five laps of a stint and a 44.6% improvement when the baseline spans a pit stop. In settled running (laps 11+), the advantage drops to 10.2% and is not statistically significant at the race level.
Ablation Findings: All three residual configurations beat the persistence baseline, while no absolute configuration does. The worst absolute model underperforms the baseline by a factor of seven. This confirms that the advantage stems from predicting the residual, not from specific feature engineering.
Degradation: The study finds that degradation models using a stint-referenced target do not provide independent evidence of degradation forecasting skill. The GRU model's apparent success on degradation is merely a reflection of its lap-time performance.
Out-of-Time Performance: The frozen model retains its advantage in 2024 (14.3%) and 2025 (9.8%), though the performance varies by season.
Training Era: Training on pre-2022 data alone yields results (11.9%) indistinguishable from same-era training (12.2%). Pooling pre- and post-2022 data yields the best performance (16.1%), suggesting historical data is beneficial rather than harmful.

Significance and Claims
The paper claims that the credibility of race-strategy decision support is bounded by the credibility of its underlying forecasts. By applying established evaluation standards from general time-series forecasting (specifically the use of simple baselines and scaled error metrics) to Formula 1, the study reveals that:

Reporting a simple baseline is necessary but insufficient; the baseline must be constructed fairly (e.g., accounting for tyre compounds) to avoid inflating model performance.
Target formulation dictates success: The residual formulation is the primary driver of performance, not the complexity of the feature set or the model architecture.
Degradation targets require re-evaluation: One-step degradation targets that subtract a within-stint constant are structurally incapable of testing degradation knowledge separately from lap-time knowledge.
Historical data is valuable: Contrary to the intuition that models must be re-fitted immediately after regulation changes, accumulating data across eras improves predictive performance.

The authors conclude that while gradient-boosted models do outperform simple forecasts, the practical value is concentrated in specific strategic windows (pit stops and stint starts). They caution against interpreting one-step degradation results as evidence of tyre-specific learning and emphasize that future work must utilize multi-lap horizons or stint-level integrals to genuinely test degradation forecasting.

Technical Summary: Baseline-Relative Evaluation of Lap-Time and Tyre-Degradation Models for Formula 1 Race Strategy

More like this