Model comparison and reproducible regression checks

Updated replication workflow

The replication code has been updated because the apc package used by the original implementation has been removed from CRAN. The revised code no longer depends on apc and uses the current clmplus implementation instead. These changes are intended to preserve the numerical results of the original analysis while ensuring that the replication workflow can be run using packages that are currently available.

The strings "a", "ac", "ap", and "apc" below are model identifiers, not package dependencies.

Deterministic comparison

data("AutoBI", package = "ChainLadder")
triangle <- ChainLadder::incr2cum(AutoBI$AutoBIPaid)
prepared <- AggregateDataPP(triangle)
model_ids <- c("a", "ac", "ap", "apc")
fits <- setNames(lapply(model_ids, function(id) {
  clmplus(prepared, hazard.model = id, verbose = FALSE)
}), model_ids)
predictions <- lapply(fits, predict)

The age-only fit reproduces chain ladder.

mack <- ChainLadder::MackChainLadder(triangle)
age_prediction <- predictions[["a"]]
stopifnot(isTRUE(all.equal(
  unname(age_prediction$full_triangle),
  unname(mack$FullTriangle),
  tolerance = 1e-10
)))
data.frame(
  accident_period = seq_along(age_prediction$reserve) - 1L,
  reserve = unname(age_prediction$reserve),
  ultimate = unname(age_prediction$ultimate_cost)
)
#>   accident_period   reserve  ultimate
#> 1               0      0.00  63918.00
#> 2               1  11995.02  74756.02
#> 3               2  28848.23  89937.23
#> 4               3  47752.62  99797.62
#> 5               4  67734.68 107290.68
#> 6               5  74742.25  96776.25
#> 7               6  97541.29 109482.29
#> 8               7 102444.81 105245.81
sum(age_prediction$reserve)
#> [1] 431058.9

Partial forecasting fills only the requested future calendar diagonals.

one_year <- predict(fits[["apc"]], forecasting_horizon = 1)
dim(one_year$full_triangle)
#> [1] 8 8
dim(one_year$apc_output$lower_triangle_apc)
#> [1] 8 1
plot(fits[["apc"]])

plot(predictions[["apc"]])