Captures detect_regime() arguments without running detection.
Returns a closure that the consumer (fit_* or backtest()) invokes
on its own internal triangle. The point is conditional /
deferred detection – the change points depend on which cells the
caller decides to expose:
In
fit_ratio()/fit_loss()/fit_premium(), the spec is invoked on the full triangle the user passed in.In
backtest(), the spec is invoked on the masked triangle of each holdout fold, never on the full triangle. Held-out diagonals are removed beforedetect_regime()sees the data, so the detected change points depend only on cells the masked fit can also see. This is the leakage-safe contract ofregime_spec().
Contrast with regime_at(), which produces an eager "Regime"
object whose change points are fixed at construction time
(independent of the fold's masked data).
Use regime_spec() when you want change points to be re-detected
per fold so backtest honestly answers "given the data available at
this fold, what regime structure would I have picked?" Use
regime_at() when you want a fixed regime tested across folds.
Arguments
- ...
kwargs passed verbatim to
detect_regime()when the spec is invoked (e.g.loss,by,min_run,method).
Value
A function of one argument (a "Triangle") returning a
"Regime" object. The caller decides which triangle to pass (full
vs. masked); inside backtest() this is always the masked training
triangle.
Examples
if (FALSE) { # \dontrun{
# Capture detection arguments, defer execution until fit time.
spec <- regime_spec(loss = "loss_ata")
# In fit_ratio(): closure is invoked on the user's `tri`.
fit <- fit_ratio(tri, loss_regime = regime_spec(loss = "loss_ata"))
# In backtest(): closure is invoked on the *masked* triangle of
# each holdout fold, so detected change points never peek at
# held-out cells.
bt <- backtest(tri, holdout = 6L,
loss_regime = regime_spec(loss = "loss_ata"))
} # }