mart_corner_skill_driver

Driver corner skill vs a LORO (leave-one-race-out) same-car baseline.

Part of the Marts family.

Driver corner skill vs a LORO (leave-one-race-out) same-car baseline. Grain: (race_year, driver_id). For each push-lap corner cell, subtracts the mean residual of the driver’s OTHER same-car teammates that race, so a driver’s edge over a teammate isn’t split in half by including themselves in their own car average. Each (driver, race, corner) cell is winsorized to +/-1.0s before the season mean so one thin/outlier cell can’t dominate a season. Each phase is z-scored within season so all three contribute equally to corner_skill_index. SC / VSC / pit / inaccurate / non-push (corrected or rain-affected) laps are excluded before car averages are formed. Negative values = faster. corner_skill_index = sum of three z-scores; lower = faster. SIGN FIX (00d, 2026-09-20). Until this date the braking phase entered the index inverted: int_corner_skill_residuals built braking_loss_s as (own - field median), so positive meant braking LATER = faster, while the mid and exit phases were positive = worse, and this model summed all three z-scores and ordered ASC. The fix was made at the SOURCE (braking_loss_s is now (field - own)), not by negating the braking z here, so that every column in this model shares one convention and corner_skill_index stays the plain sum of the three published z-scores - the identity the app prints beside them. braking_skill_s, braking_skill_se_s and braking_skill_z are therefore an exact sign flip of their pre-00d values (every step from the residual to the z is linear or odd; the STDDEV divisor and the cell counts are invariant), and corner_skill_index moved by exactly -2 * the old braking_skill_z. Both the direction and the sum identity are now pinned by transform/tests/assert_corner_skill_sign_convention.sql.

Lineage

Overview

Property	Value
Layer	marts
Family	Marts
Contract enforced	No
Tags	mart
Upstream	stg_laps, int_event_corrections, int_track_evolution, int_corner_skill_residuals
Model-level tests	unique_combination_of_columns

25 tests guard this model - 22 generic, 3 singular.

Columns

Column	Type	Nullable	Tests	Description
race_year		Yes	not_null	Season year
driver_id		Yes	not_null	FastF1 three-letter driver code (e.g. VER, HAM)
constructor_id		Yes	not_null	Driver’s team for the season (ANY_VALUE may vary for mid-season changes)
braking_skill_s		Yes	not_null, expect_column_values_to_be_between	Mean winsorized LORO-adjusted braking-phase residual (seconds). Negative = faster than the equal-car benchmark, i.e. braking LATER than the same-car reference. Bounded to [-1, 1] by the cell winsorization (convexity of the mean). 00d flipped this column’s sign at the source; a pre-2026-09-20 value of +0.0745 for VER 2024 is now -0.0745.
mid_corner_skill_s		Yes	not_null, expect_column_values_to_be_between	Mean winsorized LORO-adjusted mid-corner residual (seconds). Negative = faster.
exit_skill_s		Yes	not_null, expect_column_values_to_be_between	Mean winsorized LORO-adjusted exit-phase residual (seconds). Negative = faster.
braking_skill_se_s		Yes	not_null	Standard error of braking_skill_s (STDDEV of winsorized cells / SQRT(braking_cells_n)).
mid_corner_skill_se_s		Yes	not_null	Standard error of mid_corner_skill_s.
exit_skill_se_s		Yes	not_null	Standard error of exit_skill_s.
braking_skill_z		Yes	not_null	braking_skill_s standardised within season (zero mean, unit variance). Negative = faster on entry than the equal-car benchmark, the same direction as mid_corner_skill_z and exit_skill_z — that agreement is what 00d fixed, and it is enforced upstream at braking_loss_s rather than by a negation here. NULL below PHASE_MIN_CELLS=30 braking_cells_n — the mean is still published, only the standardized score is withheld as unreliable.
mid_corner_skill_z		Yes	not_null	mid_corner_skill_s standardised within season. NULL below 30 mid_cells_n.
exit_skill_z		Yes	not_null	exit_skill_s standardised within season. NULL below 30 exit_cells_n.
corner_skill_index		Yes	not_null, expect_column_values_to_be_between	Plain sum of the three published z-scores, all three signed the same way (negative = faster than the equal-car benchmark for that phase), so LOWER = faster driver in corner skill. No phase is negated or reweighted at summation — 00d ruled the braking sign be fixed at the source instead, precisely so this stays the sum the app prints beside the three columns; assert_corner_skill_sign_convention.sql pins it. NULL if any one phase is withheld by its cell floor (see braking_skill_z). Mid-season replacements with few races can reach ±9 because the car baseline is dominated by the full-season driver; this is expected behaviour.
braking_cells_n		Yes	not_null	Count of (race, corner) cells behind braking_skill_s this season.
mid_cells_n		Yes	not_null	Count of (race, corner) cells behind mid_corner_skill_s this season. Equal to mapped_corners: every mapped corner has a mid-apex, the most complete of the three phases.
exit_cells_n		Yes	not_null	Count of (race, corner) cells behind exit_skill_s this season. Can be much lower than mapped_corners — not every corner has a clean exit-measurement point before the next braking zone.
mapped_corners		Yes	not_null	Number of corner observations used in the season aggregate (≥ 100).