Modelling limitations
1. Sequential residualisation propagates error
The estimation strategy (see Methodology) subtracts each term in order and operates on the residual. Estimation error in an early term propagates into every downstream term, including the driver residual. If the fuel weight penalty is miscalibrated for a circuit with few qualifying stints, the compound and driver estimates absorb that error. The CI invariant proves the arithmetic closes exactly. It cannot prove the coefficients are physically correct. Trust driver residuals most at circuits where all upstream terms are well-calibrated, and treat them with more caution where any upstream term carries high uncertainty.2. Tyre coefficients need sufficient data
The Kaplan-Meier survival estimates and hockey-stick polynomial coefficients are estimated per(circuit, compound, season). Groups with fewer than approximately 8 qualifying stints produce high-variance estimates. Currently 1 of 44 circuits carries a REVIEW_REQUIRED flag on weight penalty calibration: spanish_grand_prix, in the calibration_flag column of the circuit_reference seed (transform/seeds/circuit_reference.csv, one row per circuit_key). The other 43 are SKIPPED (not calibrated, so they keep the prior). The dim_circuits table does not carry the flag.
The dim_compounds_season seed covers every circuit × compound × season group. Not all are equally well-calibrated.
3. No clean holdout season yet
2025 is ingested and decomposed, and the models train on 2018–2025. The designated holdout is 2026, which has no data yet. Until it does, ML evaluation rests on the final cross-validation fold (2025), which was also used to admit features, so it is a validation fold rather than a clean holdout.4. No 2018 Rounds 1–2 telemetry
F1’s archive is missing the race telemetry streams for 2018 rounds 1 and 2: Australia has the car-data stream but no position stream (FastF1 needs both to merge per-lap telemetry), and Bahrain has neither. Those two races (Australia and Bahrain) have no telemetry data in the Bronze layer. Without a gap to the car ahead, no car except the leader is known to be in clean air, and a missing measurement is recorded as unknown, not as clear air. Every lap of both races therefore falls short of the five clean-air cars a field-base cell needs, so both races are excluded from the field base. They have no residual, no driver-isolation laps and no training labels; those labels are lost for that reason, not guessed. This does not affect decompositions for other races, but it means 2018 is underrepresented in coefficient estimation and in the training set.5. Weather granularity is session-level
Track temperature and weather data are sourced at session resolution. Within-race variation afternoon temperature rises, brief cloud cover is captured only coarsely by the ambient component. Compounds with high temperature sensitivity at circuits with strong afternoon temperature gradients (Bahrain, Abu Dhabi) may carry less precise decompositions as a result. Higher-resolution ambient modelling is possible with OpenF1’s 1 Hz weather data but is not yet implemented.6. Constructor coefficients are pre-season priors
Constructor structural pace priors are estimated from the ingested seasons and held fixed until the next refit. A team that undergoes a major performance shift mid-season a large floor upgrade or a regulatory regression will have a stale constructor coefficient. The driver residual absorbs that difference until the constructor model is refitted.7. ML cliff prediction is v1
The machine layer (see Tyre Cliff for how it works) is built and beats a strong per-cohort baseline across all three target families. The honest remaining constraint is narrower: the cliff classifier is the weakest model, with a macro-F1 of approximately 0.40 on the 4-class cliff timing problem. It decisively outperforms the majority prior, but absolute skill on the rare imminent-cliff windows (0_to_2 laps) is modest. There is also no live 2025 holdout yet headline numbers come from time-series cross-validation with the final fold at 2024. They will convert to a true out-of-sample reveal once 2025 data is ingested.
Data ceiling
The limitations above are modelling choices improvable with better calibration or additional code. The limits below are fundamental: they follow from what the FIA and FastF1 expose, regardless of how much modelling effort goes in. State them plainly; the project’s credibility comes from modelling around them honestly.
None of these limits invalidate the decomposition framework. They define where current estimates should be treated with more caution, where data improvements would help most, and what the roadmap is addressing next.
Additive Identity
The additive equation and the CI invariant that keeps it honest
Methodology
How sequential residualisation separates the physics terms
Residual Decomposition
Why error propagation is contained to arithmetic, not coefficient correctness