trailing_median

Median of value_col over observations STRICTLY BEFORE the current row.

Median of value_col over observations STRICTLY BEFORE the current row. The canonical point-in-time baseline. Two models compute one (int_lap_thermal_proxy’s stint baseline, 08e; int_corner_skill_residuals’ corner field median, 02g). A third point-in-time rebuild, int_sc_hazard_history (02d), needs a trailing RATE rather than a median and so calls trailing_sum() below over the same frame. Both leaks that opened 08 were a baseline built with a GROUP BY over a block that included the row it scored — write the window once, here, so the next one is a call rather than a rediscovery. Returns NULL where fewer than min_observations valid observations precede the row, rather than a median over one or two laps that is mostly noise. Those NULLs are NOT declarable from the feature contract: they are deterministic on the count of valid prior observations, which is not a contract axis. Emit trailing_observation_count() alongside as a companion column so a consumer can condition on it.

Signature

{{ trailing_median(value_col, partition_by, order_by, lookback=none, frame='rows', min_observations=1, valid_condition=none) }}

Arguments

Argument	Description
value_col	column or expression to take the median of
partition_by	list of columns, or a raw string — the entity the baseline is for
order_by	list of columns, or a raw string — the time axis
lookback	integer N for a trailing-N window, or none for expanding. Expanding pools the whole history; trailing-N keeps the window local to current conditions. The choice belongs to the caller and is a decision its leaf doc must record.
frame	’rows’ (default) or ‘range’ — see trailing_window_frame
min_observations	floor below which the result is NULL (default 1)
valid_condition	boolean expression; rows failing it contribute no observation

Returns

scalar expression composable in a SELECT list.

Usage

SELECT
      {{ trailing_median('lap_time_s', ['stint_id'], ['lap_in_stint'],
                         valid_condition='is_valid_lap') }} AS stint_baseline_pace,
      {{ trailing_observation_count('lap_time_s', ['stint_id'], ['lap_in_stint'],
                         valid_condition='is_valid_lap') }} AS baseline_observations_n
    FROM combined