TyreLife
Clear target separation makes tyre age one of the first raw variables worth testing.
ORIGINAL NOTEBOOK · WEB REPORT
DETAILED EXPLORATORY DATA ANALYSIS
Lap-by-lap, from raw data to a baseline.
A native web translation of the project notebook—rebuilt from its recorded aggregates, model outputs, section order, and conclusions.
439,140Train rows
188,165Test rows
16 / 15Train / test columns
0Missing values
53.6 MBTraining memory
TARGET DISTRIBUTION
One in five laps is followed by a pit stop.
The notebook begins with the target itself. PitNextLap is imbalanced, but not severely enough to demand aggressive resampling—especially because ROC AUC measures ranking quality.
Exact notebook data
HTML · CSSPitNextLap distribution
19.9%pit next lap
NUMERIC FEATURE DISTRIBUTIONS
Train and test overlap across the feature space.
The notebook compares nine shared numeric features. The native profiles below use its exact train and test means and standard deviations to show their overlap without relying on a raster image.
Exact notebook data
HTML · CSSTrain vs test statistical profiles
TrainTestDot = mean · band = ±1 standard deviation
ACTIVE CHANNELTyreLife
- Train μ
- 14.16
- Test μ
- 14.16
- Mean drift
- +0.02%
Main observation
Train and test distributions heavily overlap for most features. Local validation should therefore be a useful leaderboard proxy.
FEATURE VS TARGET
The separation appears in tyre life and race timing.
The notebook's class overlays point to tyre life and race timing. Because their histogram bins were not saved in the notebook JSON, the native chart uses the exact printed target correlations as a reproducible signal summary.
Exact notebook data
HTML · CSSFeature signal toward PitNextLap
LapNumber
Later race laps carry more pit-stop likelihood, although timing features overlap.
RaceProgress
Strategy phases emerge more clearly than they do in raw LapTime_Delta.
CORRELATION HEATMAP
Strong signal, redundancy, and one inverted feature.
The relationship map preserves the notebook's most consequential feature-to-feature and feature-to-target values in a responsive, readable form.
Exact notebook data
HTML · CSSRelationship map
+0.27TyreLife
+0.27LapNumber
+0.20Stint
+0.19RaceProgress
-0.17Cumulative_Degradation
-0.005LapTime_Delta
Redundancy warning
LapNumber and RaceProgress correlate at 0.96; TyreLife and LapNumber at 0.65. Interaction features may be more useful than carrying every raw timing signal independently.
CATEGORICAL ANALYSIS
The compound ordering defies real-world intuition.
The notebook analyses pit rate by compound and year. These exact group means are now rendered as native bars, keeping the 2023 anomaly visible without the old Matplotlib panel.
Exact notebook data
HTML · CSSCategorical pit rates
HARD 32.8%SOFT 19.3%INTERMEDIATE 15.2%MEDIUM 10.1%WET 2.5%
Notebook warning
HARD has the highest pit rate—not SOFT. This synthetic dataset does not fully honour domain logic, so assumptions must not override the observed data.
PIT TIMING
There is no single universal pit window.
The notebook answers when stops happen by race lap, race progress, tyre age, and compound. Its raw histogram bins were not serialized, so this native view clearly labels the observed ranges as notebook-derived rather than exact bin counts.
Notebook-derived
HTML · CSSObserved pit windows
LapNumber
Stops spread broadly across laps 5-55, consistent with mixed one-stop and two-stop strategies.
RaceProgress
Activity peaks around 0.35-0.60 and tapers sharply after 0.70.
TyreLife
The strongest concentration is 10-20 laps; stops become rare after 30 laps.
By compound
Tyre-life distributions overlap heavily, so compound alone does not determine stint length.
DEGRADATION ANALYSIS
LapTime_Delta is inverted, not useless.
The notebook bins tyre age and prints the average LapTime_Delta for every interval. Those exact grouped means drive the native chart below.
Exact notebook data
HTML · CSSLapTime_Delta by tyre-life bin
Tyre life in laps
0-5 laps-7.41s5-10 laps-3.52s10-15 laps-2.93s15-20 laps-2.69s20-25 laps-2.58s25-30 laps-2.18s30-35 laps-1.56s35-40 laps-1.44s
Feature-engineering direction
The raw delta is relative to a reference pace. Rolling rate-of-change should be more useful than interpreting it as direct lap-over-lap tyre degradation.
TRAIN / TEST DRIFT
Only two features move by roughly five percent.
The notebook computes mean differences feature by feature. Every numeric feature remains under ten percent drift, and most stay below half a percent.
Exact notebook data
HTML · CSSTrain vs test mean drift
FeatureMean difference
TyreLife+0.02%
LapNumber-0.24%
Position-0.27%
LapTime (s)+0.04%
LapTime_Delta+5.10%
Cumulative_Degradation-0.50%
RaceProgress-0.29%
Position_Change+5.18%
Stint-0.27%
Decision
No drift correction was required. Position_Change reaches 5.18% and LapTime_Delta 5.10%; all other mean differences are below 0.5%.
RACE-LEVEL PIT RATE
Circuit identity carries real strategy signal.
The notebook ranks every race by observed PitNextLap rate. Its printed top and bottom ten rows are recreated below, including Pre-Season Testing.
Exact notebook data
HTML · CSSPit rate by race
SELECTED CIRCUITChinese Grand Prix
Pit rate38.9%
Observed laps7,311
Approx. positives2,841
HighestChinese Grand Prix38.9%
HighestMonaco Grand Prix35.7%
HighestSpanish Grand Prix32.0%
LowestItalian Grand Prix13.2%
LowestMiami Grand Prix10.4%
LowestMexico City Grand Prix9.1%
Recommendation from the notebook: encode Race because circuit-level pit rates vary substantially. Remove Pre-Season Testing because its stops do not follow race-strategy logic.
BASELINE LIGHTGBM
EDA decisions flow into a time-aware baseline.
The notebook closes by encoding Driver, Compound, and Race; holding out 2025; training LightGBM; evaluating AUC; inspecting feature importance; and writing a submission.
01Encode
Driver · Compound · Race
02Train
346,246 rows · 2022-2024
03Validate
92,894 rows · 2025
04Score
ROC AUC · year-aware
0.89754Validation AUC
227Best iteration
0.284Validation positive rate
13Model features
Exact notebook data
HTML · CSSLightGBM feature importance
Submission rows188,165
Prediction mean0.3037
Prediction std0.3469
Prediction range0.0002-0.9931
NOTEBOOK CONCLUSION
The data is clean. The assumptions are the risky part.
HighRemove Pre-Season Testing rowsHighUse a year-based validation splitMediumEncode Compound and RaceMediumEngineer rolling degradation signalsLowSkip drift correction