Skip to content

[FEAT] Mack prediction error for a portfolio of correlated triangles, after Braun 2004 #1671

Open

Description

@scks1

Description

MackChainladder gives a prediction error for each index row, but nothing in chainladder gives one for the sum of several rows. Adding the rows in quadrature treats the lines as independent. Adding the standard errors treats them as perfectly correlated. Running Mack on the summed triangle changes the reserve itself. Braun 2004 extends Mack's recursions to several triangles whose development factors correlate within each development period. His worked example uses the data in cl.load_sample("liab"), and he prints every intermediate value.

Is your feature request aligned with the scope of the package?

  • Yes, absolutely!
  • No, but it's still worth discussing.
  • N/A (this request is not a codebase enhancement).

Describe the solution you'd like, or your current workaround.

A fitted MackChainladder on a triangle with several index rows could expose a portfolio prediction error next to total_mack_std_err_. A standalone function that takes the fitted model would also work. Braun needs one extra estimate per development period and pair of rows. His equation 26 gives it, and the rest reuses the sigma, factor and projection arrays that Mack already builds. I can follow whatever shape you prefer.

Do you have any additional supporting notes?

Braun, C. 2004. The Prediction Error of the Chain Ladder Method Applied to Correlated Run-off Triangles. ASTIN Bulletin 34(2), 399-423. https://doi.org/10.2143/AST.34.2.505150

His Tables 1 and 2 match liab cell for cell. liab labels the origins 2000 to 2013 where Braun uses 1987 to 2000. A short numpy prototype of his recursions reproduces his portfolio figure on liab. The same prototype also runs on clrd paid losses summed by line of business.

Braun 2004 chainladder or prototype
liab IBNR, General and Auto 6,155,261 and 2,063,612 6,155,261 and 2,063,612
liab Mack prediction error, General and Auto 427,289 and 162,872 427,289 and 162,872
liab portfolio with no correlation 457,278 457,278
liab portfolio, Braun 509,075 509,075
clrd six lines, no correlation 752,644
clrd six lines, Braun 886,262
clrd six lines, perfect correlation 1,290,952

These figures need Development(sigma_interpolation="mack"). The default log-linear fill gives 428,317 and 163,151 for the two liab lines. Braun prints no correlation for the last development period, and setting it to zero reproduces his tables. The prototype handles more than two triangles by summing Braun's pairwise terms.

chainladder values
import numpy as np
import chainladder as cl

liab = cl.load_sample("liab")
mk = cl.MackChainladder().fit(cl.Development(sigma_interpolation="mack").fit_transform(liab))
print("liab IBNR", mk.ibnr_.sum("origin").values.ravel().round(0))
se = mk.total_mack_std_err_.values.ravel()
print("liab Mack se", se.round(0))
print("liab independent", np.sqrt((se ** 2).sum()).round(0))
mk = cl.MackChainladder().fit(liab)
print("liab Mack se, default fill", mk.total_mack_std_err_.values.ravel().round(0))

clrd = cl.load_sample("clrd").groupby("LOB").sum()["CumPaidLoss"]
mk = cl.MackChainladder().fit(cl.Development(sigma_interpolation="mack").fit_transform(clrd))
se = mk.total_mack_std_err_.values.ravel()
print("clrd LOB", list(clrd.index.iloc[:, 0]))
print("clrd Mack se", se.round(0))
print("clrd independent", np.sqrt((se ** 2).sum()).round(0))
print("clrd comonotonic", se.sum().round(0))
Braun prototype
import numpy as np
import chainladder as cl


def params(X):
    n = X.shape[0]
    f = np.zeros(n - 1)
    s2 = np.zeros(n - 1)
    for t in range(n - 1):
        r = np.arange(n - 1 - t)
        f[t] = X[r, t + 1].sum() / X[r, t].sum()
        if len(r) > 1:
            s2[t] = (X[r, t] * (X[r, t + 1] / X[r, t] - f[t]) ** 2).sum() / (len(r) - 1)
    s2[-1] = min(s2[-2] ** 2 / s2[-3], s2[-3])
    return f, s2


def rho_pair(C, D, f, g):
    n = C.shape[0]
    rho = np.zeros(n - 1)
    for t in range(n - 2):
        r = np.arange(n - 1 - t)
        sq = np.sqrt(C[r, t] * D[r, t])
        w2 = sq.sum() ** 2 / (C[r, t].sum() * D[r, t].sum())
        e = (C[r, t + 1] / C[r, t] - f[t]) * (D[r, t + 1] / D[r, t] - g[t])
        rho[t] = (sq * e).sum() / (len(r) - 2 + w2)
    return rho


def project(X, f):
    n = X.shape[0]
    H = X.copy()
    for i in range(n):
        for t in range(n - 1 - i, n - 1):
            H[i, t + 1] = H[i, t] * f[t]
    return H


def pair_msep(C, D, f, g, rh):
    n = C.shape[0]
    Ch = project(C, f)
    Dh = project(D, g)
    cs = np.array([C[: n - 1 - t, t].sum() for t in range(n - 1)])
    ds = np.array([D[: n - 1 - t, t].sum() for t in range(n - 1)])
    sq = np.array([np.sqrt(C[: n - 1 - t, t] * D[: n - 1 - t, t]).sum() for t in range(n - 1)])
    cx = rh * sq / (cs * ds)
    rnd = np.zeros(n)
    est = np.zeros((n, n))
    for i in range(n):
        for t in range(n - 1 - i, n - 1):
            rnd[i] = rnd[i] * f[t] * g[t] + np.sqrt(Ch[i, t] * Dh[i, t]) * rh[t]
        for j in range(n):
            for t in range(n - 1 - min(i, j), n - 1):
                est[i, j] = est[i, j] * f[t] * g[t] + Ch[i, t] * Dh[j, t] * cx[t]
    return rnd.sum() + est.sum()


def braun(tris):
    P = [params(X) for X in tris]
    N = len(tris)
    M = np.zeros((N, N))
    for a in range(N):
        for b in range(a, N):
            rh = P[a][1] if a == b else rho_pair(tris[a], tris[b], P[a][0], P[b][0])
            M[a, b] = M[b, a] = pair_msep(tris[a], tris[b], P[a][0], P[b][0], rh)
    return M


liab = cl.load_sample("liab")
M = braun([liab.loc[k].values[0, 0].astype(float) for k in ("GeneralLiab", "AutoLiab")])
print("liab Braun portfolio", np.sqrt(M.sum()).round(0))

clrd = cl.load_sample("clrd").groupby("LOB").sum()["CumPaidLoss"]
M = braun([clrd.values[i, 0].astype(float) for i in range(clrd.shape[0])])
print("clrd Braun portfolio", np.sqrt(M.sum()).round(0))

#412 asked for portfolio Mack support in 2023 and closed with no comments.

Would you be willing to contribute this ticket?

  • Yes, absolutely!
  • Yes, but I would like some help.
  • No.

(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