Rows and values

filter, formula, aggregate, resample, unpivot, score, sort, limit, correlate and holdout.

filter

Keeps the rows an expression is true for.

filter: heat > threshold and community != "Kentron"

formula

Adds or replaces columns. They are computed in the order written, so one can use the one before it.

formula:
  flat_price: price_m2 * area_m2
  years_of_salary: flat_price / (wage * 12 * net_share)

aggregate

Summarises groups of rows: by names the columns that make a group, and every other key is a new column computed with an aggregate function.

aggregate:
  by: [code, community]
  heat: median(lst_c)
  hot_share: share(lst_c > 45)
  days: count()

resample

Changes the time step of each series: one row per period.

resample: { to: year, how: mean }
resample: { to: month, how: { rain: sum, temp: mean }, by: [code] }

Option	Meaning
`to`	`year`, `quarter`, `month`, `week`, `day` or `season`.
`how`	`mean` (default), `median`, `sum`, `min`, `max`, `first`, `last` or `count`; or a mapping per column.
`by`	Extra columns that keep groups apart.

The table needs a time column.

unpivot

Turns columns into rows.

unpivot: { keep: [time, code], name: variable, value: value, columns: [ndvi, rain] }

Option	Meaning
`keep`	Columns every new row keeps (required).
`name`	The column holding the old column names (default `column`).
`value`	The column holding the values (default `value`).
`columns`	The columns to turn into rows; by default every metric column.

score

A weighted composite score: each weighted column becomes a percentile rank (reversed for a negative weight), and score is their weighted mean, from 0 to 100. Adds score and rank (1 is best).

score:
  weights: { heat: 2, green_share: -1, elderly: w_elderly }
  per: [district]
  top: 5

Option	Meaning
`weights`	`column: weight`; a weight may be a parameter; a negative weight favours low values.
`per`	Rank within these groups.
`top`	Keep this many best rows per group.

sort

Orders rows by columns; -column sorts descending.

sort: [-score, name]

limit

Keeps the first rows. The number may be a parameter.

limit: 20

correlate

Correlates two columns, optionally per group. The result has the by columns, n, r and method.

correlate: { x: temp, y: pm25, method: spearman, by: [station] }

method is pearson (default) or spearman.

holdout

Checks whether a ranking holds in a later period: ranks what by names by value on the rows test is false for, and again on the rows it is true for, and compares them.

check:
  holdout: { from: cells, test: year == 2025, by: code, value: heat, top_share: 0.2 }

The result has units, train_rows, test_rows, rank_correlation, top_precision, top_jaccard and top_share.