Construct a Regime object from manually specified regime changes

User-facing helper for hand-specifying a regime change (or a set of per-group changes) without running detect_regime(). The returned "Regime" object plugs into any function that consumes a Regime – fit_ratio(), fit_loss(), fit_premium(), backtest(), and the regime-change resolver – by carrying the same $changes schema as detect_regime() output.

Argument syntax mirrors data.frame() / data.table(): named vectors of equal length, one of which must be change. Any other named arguments are treated as group columns.

Usage

regime_at(..., treatment = c("segment_bridged", "segment_bridged_borrowed"))

Arguments

...

Named vectors of equal length. Must include change (coercible to Date; the start-of-regime date for the post-change regime). Any other named arguments are interpreted as group column values (e.g. coverage, channel). With no group columns the result is a pooled (single-row) Regime.

treatment

How downstream fits should apply this Regime when $changes contains multiple change points. "segment_bridged" (default) pools the bridged development band into a single factor estimate. "segment_bridged_borrowed" estimates per-segment factors and borrows the late-dev factors a segment cannot reach. Both mask to the bridged band so every cohort projects to full development. See detect_regime() for full semantics.

Value

An object of class "Regime" with the minimal schema needed by downstream consumers:

method

"manual".

loss

NA_character_ (no detection metric).

changes

data.table with columns [<group cols>..., change, regime_id, pre_value, post_value, magnitude]. regime_id is 2L (post-change regime) for each row; the stats columns are NA_real_.

groups

Character vector of group column names (possibly empty).

multi_group

TRUE when there are group columns and more than one unique group row.

Detection-specific slots (labels, trajectory, pca, dropped, n_regimes, window, window_mode, pca) are left empty / NA so the object can still be printed and consumed but is clearly distinguishable from a detected Regime.

See also

detect_regime()

Examples

if (FALSE) { # \dontrun{
# Pooled change (no group columns)
regime_at(change = "2024-07-01")

# Single-group change
regime_at(coverage = "surgery", change = "2024-04-01")

# Multiple groups, one column
regime_at(coverage = c("surgery", "cancer"),
          change   = c("2024-04-01", "2023-09-01"))

# Multi-dimensional group keys
regime_at(coverage = c("surgery", "surgery"),
          channel  = c("online", "agent"),
          change   = c("2024-04-01", "2024-05-01"))
} # }