Detect structural regime shifts across underwriting cohorts

Detect structural change points in the sequence of cohort-level development trajectories. Each underwriting cohort (indexed by the cohort_var of a "Triangle" object) is treated as a feature vector whose entries are the selected value_var observed at development periods 1, ..., K. Cohorts are then ordered by underwriting period and tested for structural shifts in the multivariate sequence.

Three detection strategies are supported:

"ecp": Multivariate non-parametric divisive change-point detection via ecp::e.divisive(). The number of regimes is determined by the data; only significant breakpoints at sig_level are retained. Preferred when the number of regimes is not known in advance.
"pelt": Univariate mean change-point detection via changepoint::cpt.mean() with the PELT algorithm applied to the first principal component of the cohort feature matrix. Fast and may return multiple breakpoints.
"hclust": Ward hierarchical clustering on the scaled cohort feature matrix, cut to n_regimes clusters. Ignores time ordering — useful as a sanity check since non-adjacent cohorts may cluster together if the trajectory pattern is not strictly chronological.

Usage

detect_cohort_regime(
  x,
  value_var = "clr",
  K = 12L,
  method = c("ecp", "pelt", "hclust"),
  n_regimes = NULL,
  sig_level = 0.05,
  min_size = 3L,
  ...
)

# S3 method for class 'CohortRegime'
print(x, ...)

# S3 method for class 'CohortRegime'
summary(object, ...)

# S3 method for class 'summary.CohortRegime'
print(x, ...)

Arguments

x: An object of class "Triangle". Must correspond to a single group (no group_var or a single-value group_var subset). Also used by S3 print() method on CohortRegime objects.
value_var: Column name of the trajectory variable. Default is "clr" (cumulative loss ratio).
K: Integer. Common development-period window used to build the cohort feature matrix. Cohorts with fewer than K observed periods are dropped. Default is 12.
method: One of "ecp", "pelt", "hclust".
n_regimes: Integer. Number of regimes to force. NULL means auto-detect for "ecp" and "pelt"; ignored (required to equal the requested value) for "hclust", where the default is 2.
sig_level: Significance level for "ecp". Default 0.05.
min_size: Minimum segment size for "ecp". Default 3.
...: Reserved for future use.
object: An object of class "CohortRegime". Used by the S3 summary() method.

Value

An object of class "CohortRegime" with components:

call: Matched call.
method: Detection method used.
value_var, K: Trajectory variable and window.
cohort_var: Period variable from x.
labels: data.table with one row per analysed cohort: period, regime id, regime label.
breakpoints: Date vector of breakpoint dates (each is the first cohort of a new regime; excludes the initial regime start).
n_regimes: Number of regimes detected.
trajectory: Cohort feature matrix (rows = cohorts, columns = development periods 1, ..., K).
pca: prcomp object fitted to the feature matrix.
dropped: Cohorts excluded due to the K window constraint.

Examples

if (FALSE) { # \dontrun{
data(experience)
exp <- as_experience(experience)
tri_sur <- build_triangle(exp[cv_nm == "SUR"], group_var = cv_nm)

# Hierarchical clustering (no extra package dependency)
r <- detect_cohort_regime(tri_sur, K = 12, method = "hclust",
                          n_regimes = 2L)
print(r)
summary(r)
plot(r)

# ecp divisive change-point detection (requires the ecp package)
r_ecp <- detect_cohort_regime(tri_sur, K = 12, method = "ecp")
} # }

Detect structural regime shifts across underwriting cohorts

Usage

Arguments

Value

See also

Examples