fct_cliff_prediction_features

Lap-grain feature table for the tyre cliff XGBoost model.

Part of the Marts family.

For the conceptual foundation, read Fct Cliff Prediction Features.

Lap-grain feature table for the tyre cliff XGBoost model. One row per valid race lap. Target is next_lap_degradation_jump_s (NULL on the last lap of each stint). Excludes synthetic-teammate features to prevent causal label leakage into a predictive model.

Breaking a column type on this model fails the dbt build - this is the enforced handoff contract with ml/.

Lineage

Overview

Property	Value
Layer	marts
Family	Marts
Contract enforced	✅ Yes
Upstream	int_lap_residual_decomposed, int_lap_anomaly_flags, int_compound_cliff_predicted, int_lap_thermal_proxy, int_lap_air_state, int_lap_proximity, int_lap_corner_inputs, int_event_corrections, int_qualifying_driver_summary, int_sc_hazard_history, int_lap_residual_stint_detrend, int_lap_telemetry_aggregates, dim_compounds_season, dim_circuits
Model-level tests	expect_column_values_to_be_between, expect_column_values_to_be_between

3 tests guard this model - 2 generic, 1 singular.

Columns

Column	Type	Nullable	Tests	Description
lap_id	varchar	No		Surrogate PK from stg_laps
stint_id	varchar	No		Stint surrogate key
race_year	integer	No		Season year
race_id	varchar	No		Race identifier
circuit_key	varchar	No		Circuit identifier (track_id via race_to_track); L0-1 cohort key and predictions output column
driver_id	varchar	No		FastF1 three-letter driver code
constructor_id	varchar	Yes		Team name
lap_number	integer	No		Lap number within race
lap_in_stint	bigint	Yes		Lap position within current tyre stint
age_in_stint	integer	Yes		Tyre age in laps (may differ from lap_in_stint for used tyres)
compound	varchar	Yes		Tyre compound (SOFT, MEDIUM, HARD, etc.)
compound_grip_peak	double	Yes		Fitted peak grip coefficient for this compound × circuit × season
compound_wear_gradient	double	Yes		Fitted wear gradient (s/lap) for this compound × circuit × season
compound_optimal_temp_low	double	Yes		Lower bound of optimal operating temperature window (°C)
compound_optimal_temp_high	double	Yes		Upper bound of optimal operating temperature window (°C)
compound_cliff_onset_laps	double	Yes		Expected lap age at cliff onset for this compound × circuit × season
compound_cliff_severity	double	Yes		Fitted cliff severity (s/lap post-cliff) for this compound × circuit × season
push_residual	double	Yes		Lap pace against the stint’s point-in-time baseline, from int_lap_thermal_proxy. NULL until the stint has one valid prior lap (08e); the NULL is deterministic on baseline_observations_n below.
cumulative_push_load_surface	double	Yes		EW cumulative push load (τ≈3 laps) for tyre surface
cumulative_push_load_bulk	double	Yes		EW cumulative push load (τ≈5 laps) for tyre bulk
baseline_observations_n	bigint	Yes		Valid prior laps behind this stint’s baseline. NOT a feature — it is not in ml/src/schema.py’s FEATURE_COLUMNS, and putting it in X is an add-ablation someone has to gate. It ships in the contract because the four thermal columns are NULL exactly where it is 0, and unlike 02a’s rebuild that missingness is not declarable from any other contract column: it rises with SC and pit disruption, so it is plausibly correlated with the outcome.
dirty_air_share_lap	decimal(2,1)	No		Fraction of this lap spent in dirty air (sector 2)
dirty_air_thermal_load_surface	double	No		EW dirty air thermal load on tyre surface
dirty_air_thermal_load_bulk	double	No		EW dirty air thermal load on tyre bulk
air_state_dominant	varchar	No		Dominant air state this lap (free_air, tow_zone, drs_train, dirty_air)
share_lap_within_1s	double	No		Share of the lap’s 100 track bins with a car within 1 s ahead on track (position channel; mean 0.2011)
share_lap_within_2s	double	No		Share of the lap’s track bins with a car within 2 s ahead on track (mean 0.4063)
share_lap_in_train	double	No		Share of the lap in a train rather than a single tow - within 1 s of the car ahead AND that car within 1 s of its own car ahead (mean 0.0800)
share_lap_behind_within_1s	double	No		Share of the lap’s track bins with a car within 1 s BEHIND. The half of traffic DistanceToDriverAhead structurally cannot see (mean 0.2012)
time_within_1s	double	No		Seconds of the lap spent within 1 s of the car ahead, time-weighted by real bin duration rather than counted in bins (mean 18.36 s)
gap_ahead_min_s	double	Yes		Closest the car ahead on track came this lap (s). NULL = a full lap of clear air; median 0.92 s
gap_ahead_median_s	double	Yes		Median gap to the car ahead on track across the lap’s bins (s). NULL = clear air; median 2.26 s
ahead_identity_stability	double	Yes		Share of the lap’s within-3 s bins held by the single most frequent car ahead. 1.0 = a settled follow, low = the car ahead rotated. NULL on the 27.3% of laps with no car within 3 s
n_distinct_cars_ahead_3s	integer	Yes		How many distinct cars were within 3 s ahead at some point in the lap (mean 1.348)
corner_input_coverage	double	No		02c. Share of the lap’s measured corners that cleared int_corner_skill_residuals’ field_corner_sample_n >= 5 gate, so carry a residual at all (mean 0.794 over mart rows, 0.799 over the training-eligible ones; 0.813 over int_lap_corner_inputs, which is the same column before the COALESCE below adds the uncovered laps. Laps average 16.35 corners measured, 13.27 of them mapped). COALESCE-d to 0.0 on the 2.5% of mart laps with no corner row, because a lap with no mapped corner genuinely has zero coverage. NOT a nuisance column: coverage is correlated with the label (the 4.15% of training-eligible rows with no braking mean average +1.393 s of 5-lap jump against -2.330 s for covered rows, sd 10.196 vs 5.938), so it is admitted explicitly and ablated separately rather than left to XGBoost as an implicit NaN pattern.
corner_braking_loss_mean_s	double	Yes		02c. Mean braking_loss_s across the lap’s corners that have a braking point (mean +0.0065 s since 00d; -0.0065 s before). Distance-of-first-brake-application vs the backward-only trailing-5 field median (02g), converted to seconds at s2 trap speed, signed (field - own) so POSITIVE = braked earlier than the field = time lost. NULL, never 0.0, where the lap has no measured braking corner - 0.0 would assert the driver braked exactly at the field median.
corner_braking_loss_sd_s	double	Yes		02c. Sample sd of braking_loss_s across the lap’s corners (mean 0.3289 s; NULL on 2.57% of corner-covered laps, which have fewer than two measured braking corners). Unchanged by 00d’s sign fix - a sample sd is invariant to negation. The tyre-management statistic of the three: a driver even across the lap loads the tyre differently from one spiking in two corners and coasting the rest, at equal mean.
corner_braking_loss_max_s	double	Yes		02c. Largest (most positive) braking_loss_s on the lap - the corner where the driver lost the most time on entry (mean +0.6475 s). 00d: this is the one corner-input column that is NOT a sign flip of its pre-00d self, because MAX does not commute with negation. It changed from “the corner braked latest” to “the corner braked earliest relative to the field”. Pre-00d mean +0.5459 s, corr(old, new) = 0.080 over 131,482 laps. Mart-only and outside FEATURE_COLUMNS, and 02c’s arms over it are already void and must re-run.
corner_mid_residual_mean_s	double	Yes		02c. Mean mid_corner_residual_s across the lap’s corners (mean -0.0102 s). Apex-speed deficit vs the trailing-5 field median, in seconds. Highest-coverage of the three phases (81.12% of corner rows, and 100% of the mapped ones) because every corner has an apex, while not every corner has a discrete braking zone or a clean full-throttle point.
corner_mid_residual_sd_s	double	Yes		02c. Sample sd of mid_corner_residual_s across the lap’s corners (mean 0.1439 s; NULL on 2.39% of corner-covered laps).
corner_mid_residual_max_s	double	Yes		02c. Largest mid_corner_residual_s on the lap - the corner with the biggest apex-speed deficit (mean 0.2543 s).
corner_exit_residual_mean_s	double	Yes		02c. Mean exit_residual_s across the lap’s corners that reach full throttle (mean 0.0589 s). Distance-to-full-throttle vs the trailing-5 field median, in seconds. Sparsest of the three phases (47.00% of corner rows, 57.94% of the mapped ones): throttle_point_m is NULL where the driver never reaches 100% throttle before the next apex, which is a real answer about a corner sequence rather than missing data.
corner_exit_residual_sd_s	double	Yes		02c. Sample sd of exit_residual_s across the lap’s corners (mean 0.2115 s; NULL on 9.78% of corner-covered laps, the sparsest phase).
corner_exit_residual_max_s	double	Yes		02c. Largest exit_residual_s on the lap - the corner the driver was slowest back to power in (mean 0.4367 s).
quali_push_laps_n	bigint	No		Push laps this driver set across Q1/Q2/Q3 this weekend. 0, not NULL, when the driver has no row in int_qualifying_driver_summary.
quali_constructor_pace_mean_s	double	Yes		Mean constructor_component_s over the driver’s qualifying push laps this weekend - car pace at low fuel, new tyres, near-zero traffic.
quali_constructor_pace_se_mean_s	double	Yes		Mean estimation SE of constructor_component_s over the same push laps.
quali_pace_delta_best_s	double	Yes		Driver’s single best qualifying push lap this weekend (MIN of quali_pace_delta_s, which is higher = slower by construction).
quali_ratio_to_segment_best_min	double	Yes		How close the driver’s best lap in any segment came to that segment’s own fastest push lap (MIN of ratio_to_segment_best).
quali_skill_session_avg_s	double	Yes		Driver’s one-lap qualifying skill for the weekend - the residual of their best push lap per segment, averaged over segments contested.
quali_segments_contested_n	bigint	Yes		Segments (1-3) in which the driver set a best qualifying push lap this weekend.
circuit_sc_hazard_per_lap	double	Yes		02d. Per-racing-lap safety-car hazard for this circuit, empirical-Bayes shrunk, from int_sc_hazard_history joined on (circuit_key, race_year) — POINT-IN-TIME as of the start of the season, estimated from races at this venue in seasons strictly earlier. Circuit x season grain, so constant across every lap of the weekend. NULL, never 0.0, where the hazard is unknowable: season 2018 is the earliest in the warehouse, so no 2018 row has a prior season to estimate from nor a prior-season pooled rate to shrink toward. 0.0 would assert a venue never throws a safety car, which is a measurement nobody made. The NaN mask is therefore EXACTLY race_year = 2018 (14,982 of 119,822 training-eligible rows, 12.5%), and race_year is not itself in FEATURE_COLUMNS — see the leaf doc’s arm D, which isolates that channel rather than letting it ride inside the hazard block.
circuit_vsc_hazard_per_lap	double	Yes		02d. As circuit_sc_hazard_per_lap, for virtual-safety-car onsets (status_code 6). Not collinear with the SC rate — corr(sc onsets, vsc onsets) = 0.125 per race over 149 races — so the two are separate channels rather than one measured twice.
circuit_any_hazard_per_lap	double	Yes		02d. SC or VSC. EXACTLY circuit_sc_hazard_per_lap + circuit_vsc_hazard_per_lap: the onset counts add and the shrinkage priors add with them, so this carries no information the other two lack. Retained because a tree cannot form the sum itself, which makes it a basis rotation rather than a redundant column.
circuit_hazard_prior_racing_laps	bigint	Yes		02d. Racing laps of prior-season exposure the hazard estimate rests on, at this venue, seasons strictly before this one. 0 on a venue’s debut season. Never NULL — the source reports it unfloored, and 0 is the measured “no prior exposure” rather than an invented value. Carried as its own channel so the ablation can separate safety-car risk from calendar tenure, which has no SC mechanism in it (arm C).
circuit_hazard_prior_seasons_n	bigint	Yes		02d. Distinct prior seasons in which this venue appears, strictly before this one. 0 on a debut season. The companion count to circuit_hazard_prior_racing_laps and the second half of arm C.
expected_compound_pace_s	double	Yes		Model-expected compound pace contribution from int_compound_cliff_predicted
expected_degradation_rate_s_per_lap	double	Yes		Model-expected degradation rate (s/lap) at current tyre age
cliff_onset_passed	boolean	Yes		TRUE if tyre has passed its fitted cliff onset lap
laps_past_cliff	double	Yes		Laps elapsed since cliff onset (0 if not passed)
ambient_temp_delta	double	Yes		Difference between current ambient temp and optimal temp midpoint (°C)
track_energy_index	double	Yes		Track energy index from dim_circuits
circuit_abrasiveness_index	integer	Yes		Track abrasiveness index from dim_circuits
n_gear_changes	integer	Yes		Gear changes this lap (telemetry powertrain group; NULL if no telemetry)
mean_rpm	double	Yes		Mean engine RPM this lap
max_rpm	double	Yes		Max engine RPM this lap
pct_full_throttle	double	Yes		Fraction of the lap at full throttle (0–1)
pct_drs_active	double	Yes		Fraction of the lap with DRS active (0–1)
short_shift_index	double	Yes		Fraction of upshifts taken below 92% of lap-max RPM (short-shifting)
mid_corner_speed_loss_kph	double	Yes		Within-stint drift: early-stint baseline minus this lap’s mean local-minimum (corner-apex) speed. Positive = slow points getting slower vs own baseline. NULL if no telemetry.
traction_wheelspin_proxy	double	Yes		Share of corner-exit full-throttle samples with non-positive accel (wheelspin proxy, 0–1)
throttle_trace_decay	double	Yes		Within-stint drift of full-throttle fraction (early-stint baseline minus this lap)
braking_point_drift_m	double	Yes		Within-stint drift: early-stint baseline minus this lap’s mean brake-onset distance. Positive = braking earlier than own baseline. NULL if no telemetry.
lift_coast_share	double	Yes		Off-throttle, non-braking, decelerating share of the lap (0–1). CONFOUND with fuel-saving consumed alongside int_lap_fuel_state, not equated with degradation.
fuel_mass_kg	double	Yes		Estimated fuel mass on board at this lap
event_flag_any	boolean	No		TRUE if any race event contaminated this lap (SC, VSC, restart, etc.)
cliff_candidate_flag	boolean	Yes		TRUE if this lap shows a MAD-score cliff spike consistent with tyre cliff. Removed from ML feature contract by 08j (2026-09-09): ruled dead on both substrates by 08g (carried no information in any family). Retained in mart for audit and historical reference; not consumed by current ML features.
anomaly_class	varchar	Yes		Anomaly classification (clean_cliff, mistake, event_driven, conditions, normal)
is_rain_lap	boolean	Yes		TRUE if rainfall was detected on this lap
surface_bulk_ratio	double	Yes		C3 feature (42nd): ratio of surface thermal load to total thermal load (cumulative_push_load_surface / (surface + bulk)). NULL when both loads are zero. Helps the model distinguish early-stint surface-dominated warming from bulk-driven structural degradation.
survival_weight	double	Yes		C2 metadata (not a feature): IPW weight = 1 / P(stint reaches lap_in_stint | compound), clipped to [0.25, 4]. Used by train.py for quantile regressor sample weighting to correct for survivorship bias (degraded stints pitted early are under-represented at high lap_in_stint values).
next_lap_degradation_jump_detrended_s	double	Yes		C1 PRIMARY target: per-lap change in driver skill residual with per-stint linear drift removed. drift_s_per_lap (OLS slope on pre-cliff laps, Step 0 median -0.072 s/lap) is subtracted so fuel-lightening and track-evolution leaks are cancelled. Bounded [-10, 10]. NULL on last lap of each stint.
next_lap_degradation_jump_s	double	Yes		Legacy target (undetrended): change in driver skill residual on the next lap, bounded [-10, 10]. Kept alongside the detrended primary for gate comparison. NULL on last lap.
next_3_lap_cumulative_jump_s	double	Yes		Cumulative detrended degradation jump over the next 3 laps: SUM over i=1..3 of (residual(t+i) - residual(t) - i * drift_s_per_lap), i.e. the current residual is subtracted once per horizon step and per-stint drift is removed on the same schedule as next_lap_degradation_jump_detrended_s. Signed and bounded [-30, 30] (the per-lap +/-10 s bound scaled by the horizon). NULL unless laps t+1..t+3 are all present in the stint — the guard is on true lap offsets, not row offsets.
next_5_lap_cumulative_jump_s	double	Yes		As next_3_lap_cumulative_jump_s over a 5-lap horizon: SUM over i=1..5 of (residual(t+i) - residual(t) - i * drift_s_per_lap). Signed and bounded [-50, 50]. NULL unless laps t+1..t+5 are all present in the stint.
laps_until_cliff_class	varchar	Yes		Bucket: 0_to_2, 3_to_5, 6_plus, none_in_stint. Laps until the FIRST crossing of the >1.0s detrended jump threshold, scanned over every remaining lap of the stint at true lap offsets. The buckets partition the horizon exactly as their names read — 6_plus is 6-or-more, not exactly 6, and none_in_stint means no crossing anywhere in the remaining stint. NULL on the last lap of a stint, where there is no horizon to scan.
is_training_eligible	boolean	No		TRUE if this lap is suitable for model training (age_in_stint > 3, no anomaly)