Repository navigation
Description
Auditable lines such as workers' comp and general liability bill a deposit premium at inception and settle the final premium by audit after the policy expires. Cullinane, "Forecasting Audit Premiums: The Role of the Actuary" (CAS Forum, Winter 2023), builds a triangle of the ratio of audit premium to deposit plus midterm endorsement premium, by policy effective month and development month (Section 4.1). He projects the ratio with chain ladder ratios or incremental differences (4.2) and multiplies the ultimate ratio by the ultimate deposit and endorsement premium to get the ultimate audit premium (4.3). chainladder already does that projection, and deposit plus audit premium is the exposure a BF or Cape Cod method should use. The docs don't show it. #1615 shows what goes wrong when a user passes the premium triangle itself as sample_weight.
Here is a small version on raa with a made-up audit ratio triangle:
import chainladder as cl
import numpy as np
import pandas as pd
print(cl.__version__, cl.__file__)
loss = cl.load_sample("raa")
# Cumulative audit premium as a ratio to deposit plus endorsement premium.
# Interim audits and midterm cancellations give the 12-month value, most
# audits arrive by 24 months, and reopened audits return a little after that.
audit_ratio = [0.02, 0.07, 0.06, 0.06, 0.06, 0.06, 0.06, 0.06, 0.06, 0.06]
deposits = {year: 40000 * 1.03**i for i, year in enumerate(range(1981, 1991))}
rows = []
for year in deposits:
for j in range(1991 - year):
rows.append({"origin": year, "valuation": year + j, "ratio": audit_ratio[j]})
ratio = cl.Triangle(
pd.DataFrame(rows), origin="origin", valuation="valuation",
columns="ratio", cumulative=True,
)
print("audit ratio link ratios", ratio.link_ratio.to_frame().mean().round(4).tolist())
ultimate_ratio = cl.Chainladder().fit(ratio).ultimate_.to_frame().iloc[:, 0].values
deposit = np.array(list(deposits.values()))
audit = ultimate_ratio * deposit
print("expected audit premium by policy year")
for year, value in zip(deposits, audit):
print(year, f"{value:,.2f}")
def exposure(amounts):
frame = pd.DataFrame({"origin": list(deposits), "premium": amounts})
return cl.Triangle(frame, origin="origin", columns="premium", cumulative=True)
bf = cl.BornhuetterFerguson(apriori=0.6)
for name, amounts in [("deposit", deposit), ("audited", deposit + audit)]:
ibnr = float(np.nansum(bf.fit(loss, sample_weight=exposure(amounts)).ibnr_.values))
print(f"BF IBNR with {name} premium: {ibnr:,.2f}")On main a59efc9:
0.10.1 /home/hatch/workspace/oss-contributions/chainladder-next-2026-11-13/src-a59efc9d/chainladder/__init__.py
audit ratio link ratios [3.5, 0.8571, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
expected audit premium by policy year
1981 2,400.00
1982 2,472.00
1983 2,546.16
1984 2,622.54
1985 2,701.22
1986 2,782.26
1987 2,865.73
1988 2,951.70
1989 3,040.25
1990 3,131.46
BF IBNR with deposit premium: 80,307.84
BF IBNR with audited premium: 85,126.31
The 24-36 link ratio below 1 comes from reopened audits that return premium, which Cullinane sees as decreases from development month 16 on, and Development handles it. His Section 4.2 notes that incremental differences are easier to work with when the sign of the data varies along a row or column. Using the audited premium raises the BF IBNR by the same 6% as the exposure. Cullinane's Section 5 makes the same point for claim frequency, which needs audited exposures too.
A gallery notebook beside docs/gallery/plot_exposure_triangle.ipynb would cover it. It should pass the audited premium, deposit plus projected audit premium, as sample_weight, not a premium development triangle (#1615), and keep origin labels equal on both sides (#1646, PR #1648). It needs no library change, so it suits a new contributor.
(This issue was written with Árbol, an AI assistant, under my direction.)
Metadata
Metadata
Assignees
Labels
Type
Fields
Priority
Scope
Effort
Projects
- StatusShow more project fieldsReady
Activity