Skip to content

[BUG] After a bootstrap fit, full_triangle_ rebuilds future cells from the ultimate and drops the simulated process-variance path #1667

Open

Description

@scks1

After Chainladder().fit() on a BootstrapODPSample triangle, full_triangle_ does not hold the path the fit simulated. MethodBase._include_process_variance adds gamma noise to each future cell and moves ultimate_ to the noisy last column. full_triangle_ then rebuilds every future cell from that ultimate with _get_full_triangle, which spreads ultimate_ - latest along the expected development pattern. The ultimates agree, but the cell-level noise is gone.

Shapland and Leong (2010), "Bootstrap Modeling: Beyond the Basics", CAS E-Forum Fall 2010, add process variance to each future incremental value (section 3) and report the cash flow distribution by calendar year (section 6.2, Figure 6.4). A calendar-year runoff built from full_triangle_, as in the plot_ibnr_runoff gallery, understates that spread.

import numpy as np
import chainladder as cl

tri = cl.load_sample("genins")
sims = cl.BootstrapODPSample(n_sims=200, random_state=42).fit_transform(tri)
model = cl.Chainladder().fit(sims)

# The path fit simulated: the expected completion of each resampled history plus process_variance_
no_noise = sims.copy()
del no_noise._get_process_variance
path = cl.Chainladder().fit(no_noise).full_triangle_ + model.process_variance_


def calendar_2011(full):
    runoff = (full.cum_to_incr() - sims.cum_to_incr()).dev_to_val()
    col = [str(v)[:4] for v in runoff.development].index("2011")
    return np.nansum(runoff.values[:, 0, :, col], axis=1)


print("ultimates match:", np.allclose(path.values[..., -1], model.full_triangle_.values[..., -1]))
print("origin 2010, sim 0, ages 24-60, path:          ", np.round(path.values[0, 0, -1, 1:5]).tolist())
print("origin 2010, sim 0, ages 24-60, full_triangle_:", np.round(model.full_triangle_.values[0, 0, -1, 1:5]).tolist())
print("sd of calendar 2011 paid, path:          ", round(float(calendar_2011(path).std())))
print("sd of calendar 2011 paid, full_triangle_:", round(float(calendar_2011(model.full_triangle_).std())))

On main cc3335c this prints

ultimates match: True
origin 2010, sim 0, ages 24-60, path:           [1231064.0, 2056950.0, 2844350.0, 3147112.0]
origin 2010, sim 0, ages 24-60, full_triangle_: [1117686.0, 1915630.0, 2824323.0, 3328947.0]
sd of calendar 2011 paid, path:           721256
sd of calendar 2011 paid, full_triangle_: 593207

None of the interior future cells match in any of the 200 draws. process_variance_ holds the noise, but it has no docstring and the noise is measured against the completion before ultimate_ moved, so full_triangle_ + process_variance_ is not the path either.

Two options I can see.

  1. Keep the simulated path on the fitted model, for example as a full_triangle_ that returns it when process variance was drawn, or as a separate attribute.
  2. Leave the behaviour and document it. Add a docstring for process_variance_ and a note on full_triangle_ that says it is not the simulated path after a bootstrap fit.

Which do you prefer? Related: #1639 and #1652 (process variance in predict) and #1664 (comparing methods on the same draws).

(This issue was written with Árbol, an AI assistant, under my direction.)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    Priority

    None yet

    Scope

    None yet

    Effort

    None yet

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions