Hybrid filtering via regime break

Motivation

When a long-term health portfolio undergoes a rate revision, coverage restructure, or underwriting overhaul, cohorts after that event behave differently from earlier ones. Fitting chain ladder on the full triangle lets old-cohort link factors leak into the new-cohort projections, which shows up as a monotone drift across diag_summary in backtest().

The recent = N argument suppresses some of this drift, but a calendar-diagonal cut is symmetric across both axes — it discards older cohorts’ young-dev cells too, where the ED region was already stable. The natural fix is asymmetric:

Pre-maturity (ED region): horizontal cut — keep only post-break cohorts.
Post-maturity (CL region): diagonal cut — keep only the recent N calendar diagonals.

regime_break implements that split.

Two-axis asymmetry

Axis	Number of changes	Source
x (maturity, ED → CL)	exactly one per group	`fit_ata$maturity`
y (regime break)	zero or many per group	`detect_cohort_regime$breakpoints`

The maturity point $k^*$ is a single internal switch produced by find_ata_maturity(). Regime breaks are exogenous events — there can be none, one, or several. When regime_break receives multiple values, the most recent is used, since post-break statistics are most stable when the post-break window has accumulated the largest number of cohorts.

API

regime_break is a shared argument on fit_ata, fit_ed, and fit_lr. It accepts:

Input	Behaviour
`NULL` (default)	no filtering — backwards compatible
`Date` or character scalar	single break date
Date/character vector	uses the latest entry
`CohortRegime` object	output of `detect_cohort_regime()` passed in

library(lossratio)
data(experience)
exp     <- as_experience(experience)
tri_sur <- build_triangle(exp[cv_nm == "SUR"], cv_nm)

# Single break date
fit_lr(tri_sur, method = "sa", recent = 18L,
       regime_break = "2024-04-01")

# CohortRegime object directly
reg <- detect_cohort_regime(tri_sur)
fit_lr(tri_sur, method = "sa", recent = 18L, regime_break = reg)

# Vector — latest is used (= 2024-04-01)
fit_lr(tri_sur, method = "sa",
       regime_break = c("2023-06-01", "2024-04-01"))

In simple modes (fit_ata, fit_ed, or fit_lr(method ∈ {"ed","cl"})) the same argument acts as a plain cohort cut.

SA-mode hybrid behaviour

The hybrid split activates only for fit_lr(method = "sa") with both regime_break and recent:

dev ≤ $k^*$ — ED region: post-break cohorts only.
dev > $k^*$ — CL region: latest recent diagonals only (full triangle if recent = NULL).

The maturity point $k^*$ itself is found in a two-pass procedure: first on the raw triangle (so noisy post-break windows do not destabilise $k^*$ ), then the hybrid filter is applied to the actual fit using the fixed $k^*$ .

plot_triangle(type = "usage") visualises which cells each filter configuration feeds to fit_lr:

plot_triangle(tri_sur, type = "usage", holdout = 6L)                                 # full
plot_triangle(tri_sur, type = "usage", recent = 12L, holdout = 6L)                   # recent
plot_triangle(tri_sur, type = "usage", regime_break = "2024-04-01", holdout = 6L)    # break
plot_triangle(tri_sur, type = "usage", recent = 12L,
              regime_break = "2024-04-01", holdout = 6L)                             # hybrid

Cells used by each filter configuration on the SUR triangle. Blue = fit data, red = held out (last 6 calendar diagonals), light grey = excluded by the filter, white = future. Vertical dashed line marks the maturity switch

k^*

; horizontal dashed line marks the regime break cohort.

The hybrid panel shows the dev-axis split that SA mode applies: a cohort cut on the ED side and a calendar diagonal cut on the CL side, joined at $k^*$ .

Case study — SUR cohort

The bundled experience dataset embeds a synthetic 2024-04 break in the SUR coverage. Backtests on the same triangle with four variants:

reg <- detect_cohort_regime(tri_sur)

bt_full   <- backtest(tri_sur, holdout = 6L)
bt_recent <- backtest(tri_sur, holdout = 6L, recent = 18L)
bt_break  <- backtest(tri_sur, holdout = 6L,
                      regime_break = reg)
bt_hybrid <- backtest(tri_sur, holdout = 6L, recent = 18L,
                      regime_break = reg)

Reproduced from dev/regime_backtest_hybrid.R:

Variant	drift (cal30 − cal25)	overall mean
full	+4.50pp	-1.25%
recent = 18	+2.03pp	-3.45%
regime_break + recent = 18	-0.69pp	+0.03%

Two columns summarise the AEG = actual / pred − 1 (positive = under- projection) measured on the held-out diagonals:

drift (cal30 − cal25): AEG aggregated by calendar diagonal, then the (latest − earliest) difference. Captures whether the prediction error is monotonically changing across the hold-out window — the signature of a regime that the static model has not absorbed.
overall mean: cell-level mean AEG across all held-out cells — the model’s directional bias.

Drift collapses from +4.50pp under full to -0.69pp under the hybrid filter; the overall mean returns to ~0. The hybrid joins two axis cuts at $k^*$ : a cohort cut on the ED side (dev ≤ k*) and a calendar diagonal cut on the CL side (dev > k*).

Multi-group handling

detect_cohort_regime() assumes a single-group triangle. For a portfolio with multiple cv_nm groups, call it per group:

fits <- lapply(unique(exp$cv_nm), function(g) {
  tri_g <- build_triangle(exp[cv_nm == g], cv_nm)
  reg_g <- detect_cohort_regime(tri_g)
  fit_lr(tri_g, method = "sa", recent = 18L,
         regime_break = reg_g)
})

A future extension may accept regime_break = list(SUR = "2024-04-01", CAN = "2023-12-01"). Today only scalar / vector / CohortRegime are supported.

Limitations

If the post-break window is too short (small n_post), the ED intensity $g_k$ and link factors $f_k$ become noisy. A practical threshold is n_post ≳ 6. Below that, prefer recent alone, or wait for credibility-weighted blending of pre- and post-break factors (planned).

Note also that regime_break only filters the data feeding link factor estimation. Once the factors are fixed, all cohorts share them, so pre-break ultimates inherit the post-break dynamics.