Repository navigation
Description
Mario Wüthrich's paper "Exponential Seniority Weighting Chain-Ladder Reserving", arXiv:2611.06126 from November 2026, derives chain ladder factors that give older link ratios exponentially less weight. Eq. (1.1) weights each ratio C_{i,j+1}/C_{i,j} by q^k · C_{i,j}, where k counts diagonals back from the latest one and q lies in (0, 1]. With q = 1 the factors equal Mack's.
Users who want recent experience to count more reach for n_periods today. It keeps the latest diagonals at full weight and drops the rest. The paper argues against a hard window because it throws away the older data, and its weights fade that data out instead. Development has no option for these weights.
The snippet below computes the factors outside the library and passes them to DevelopmentConstant. On the paper's Table 5 triangle it prints a total reserve of 6,047,064 at q = 1 and 5,989,679 at q = 0.9. Both totals match the paper's Table 2.
A Development parameter such as decay=q could multiply the existing link-ratio weights by q^k. The factors would then match the paper, but MackChainladder would still compute Mack's sigma. Its standard errors would not match the paper's RMSEP of 482,923 at q = 0.9, which comes from the paper's own state-space model. A first version could leave that part out.
Would you want this as a Development parameter, or as a gallery example?
Snippet
import numpy as np
import pandas as pd
import chainladder as cl
# Wuthrich (2026), arXiv:2611.06126, Table 5 (Wuthrich-Merz Table 1.4)
rows = [
[5946975, 9668212, 10563929, 10771690, 10978394, 11040518, 11106331, 11121181, 11132310, 11148124],
[6346756, 9593162, 10316383, 10468180, 10536004, 10572608, 10625360, 10636546, 10648192],
[6269090, 9245313, 10092366, 10355134, 10507837, 10573282, 10626827, 10635751],
[5863015, 8546239, 9268771, 9459424, 9592399, 9680740, 9724068],
[5778885, 8524114, 9178009, 9451404, 9681692, 9786916],
[6184793, 9013132, 9585897, 9830796, 9935753],
[5600184, 8493391, 9056505, 9282022],
[5288066, 7728169, 8256211],
[5290793, 7648729],
[5675568],
]
df = pd.DataFrame(
[(2000 + i, 2000 + i + j, v) for i, r in enumerate(rows) for j, v in enumerate(r)],
columns=["origin", "valuation", "paid"],
)
tri = cl.Triangle(df, origin="origin", valuation="valuation", columns="paid", cumulative=True)
C = tri.values[0, 0]
def seniority_ldf(C, q):
"""Eq. (1.1): row i at lag j gets weight q**k * C[i, j], k = diagonals back from the latest."""
f = []
for j in range(C.shape[1] - 1):
obs = [i for i in range(C.shape[0]) if not np.isnan(C[i, j + 1])]
w = np.array([q ** (obs[-1] - i) * C[i, j] for i in obs])
r = np.array([C[i, j + 1] / C[i, j] for i in obs])
f.append((w * r).sum() / w.sum())
return np.array(f)
print("q=1 matches Development():", np.allclose(seniority_ldf(C, 1.0), cl.Development().fit(tri).ldf_.values[0, 0, 0]))
for q in (1.0, 0.9):
dev = cl.DevelopmentConstant(
patterns=dict(zip(tri.development[:-1], seniority_ldf(C, q))), style="ldf"
).fit_transform(tri)
print(f"q={q}: total reserve {cl.Chainladder().fit(dev).ibnr_.sum():,.0f}")Output on main at 687e329:
q=1 matches Development(): True
q=1.0: total reserve 6,047,064
q=0.9: total reserve 5,989,679
(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 fieldsNot Ready
Activity