关于「兑付生存曲线:删失、乘积限估计与到期日外推」,《积分负债与沉睡精算》给出的定量结论是什么?

结论:第二章用 Kaplan–Meier 曲线处理删失,并显示同一份数据外推到到期日,几何、Weibull 与混合模型给出 0.083、0.179 与 0.250 的沉睡率,真实值 0.286。

数据口径:「忽略沉睡会错估负债」须分方向:完全忽略沉睡在示意数据中把负债多记 0.473,而用未成熟数据直接计数则把每分递延额少记 0.272,后者更危险。 书中参数为示意;全部计算由书中程序生成;本书是非个人化会计、税务或法律建议,以执业会计师意见为准。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.083, 0.179, 0.250, 0.286

model params loglik AIC S(24) S(36)=breakage
geometric1 -5613.5 11229.1 0.190 0.083
weibull2 -5499.5 11002.9 0.276 0.179
mixture3 -5493.4 10992.8 0.305 0.250
mixture fit: k 0.798 lam 9.70 pi 0.796
true S(36)0.286
查看完整验算程序
# Chapter 2. p22: three redemption curves fitted to the same censored data.
# Same synthetic lots as p21 (seed 20261113). Discrete-time likelihood
# on the monthly life table: sum of d log h + (n - d) log(1 - h), with
# h(t) = 1 - S(t) / S(t - 1). Models: geometric S = (1 - h) ** t;
# Weibull S = exp(-(t / lam) ** k); Weibull with a never-redeem share
# S = 1 - pi * (1 - exp(-(t / lam) ** k)). Grid search refined three
# times. Ultimate breakage is S(36), the share unredeemed at expiry.
import math
import random

N, PI, K, SCALE, HORIZON, EXPIRY = 3000, 0.75, 0.8, 9.0, 24, 36


def lots(rng):
    out = []
    for _ in range(N):
        follow = HORIZON - rng.randrange(HORIZON)
        m = None
        if rng.random() < PI:
            m = max(1, math.ceil(rng.weibullvariate(SCALE, K)))
        out.append((follow, m))
    return out


def table(data):
    tab = []
    for t in range(1, HORIZON + 1):
        n = len([1 for f, m in data if f >= t and (m is None or m >= t)])
        d = len([1 for f, m in data if m == t and f >= t])
        tab.append((t, n, d))
    return tab


def surv(model, p, t):
    if model == "geometric":
        return (1 - p[0]) ** t
    pi = p[2] if model == "mixture" else 1.0
    return 1 - pi * (1 - math.exp(-(t / p[1]) ** p[0]))


def loglik(tab, model, p):
    terms = []
    for t, n, d in tab:
        h = 1 - surv(model, p, t) / surv(model, p, t - 1)
        h = min(max(h, 1e-12), 1 - 1e-12)
        terms.append(d * math.log(h) + (n - d) * math.log(1 - h))
    return math.fsum(terms)


def grid(tab, model, axes):
    best = None
    for _ in range(4):
        pts = [[c + s * j for j in range(-6, 7)] for c, s in axes]
        cand = [(a, b, c) for a in pts[0] for b in pts[1] for c in pts[2]]
        for p in cand:
            if min(p) <= 0 or (model == "mixture" and p[2] > 1):
                continue
            ll = loglik(tab, model, p)
            if best is None or ll > best[0]:
                best = (ll, p)
        axes = [(best[1][i], s / 4) for i, (c, s) in enumerate(axes)]
    return best


if __name__ == "__main__":
    tab = table(lots(random.Random(20261113)))
    fits = (("geometric", [(0.07, 0.01), (1.0, 0.0), (1.0, 0.0)], 1),
            ("weibull", [(0.8, 0.1), (12.0, 2.0), (1.0, 0.0)], 2),
            ("mixture", [(0.8, 0.1), (9.0, 1.5), (0.75, 0.04)], 3))
    print("model      params  loglik     AIC   S(24)  S(36)=breakage")
    for name, axes, k in fits:
        ll, p = grid(tab, name, axes)
        print("{:9s} {:7d} {:8.1f} {:7.1f} {:7.3f} {:9.3f}".format(
            name, k, ll, 2 * k - 2 * ll, surv(name, p, 24),
            surv(name, p, EXPIRY)))
    print("mixture fit: k {:.3f}  lam {:.2f}  pi {:.3f}".format(*p))
    print("true S(36) = {:.3f}".format(1 - PI * (1 - math.exp(
        -(EXPIRY / SCALE) ** K))))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《会员积分负债与沉睡率精算、积分发行兑付率与递延收入的破产边界母典》第 2 章「兑付生存曲线:删失、乘积限估计与到期日外推」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-315 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026315,
  author    = {Fan, Yafu},
  title     = {Loyalty-Point Liabilities and Breakage Actuarial Science: Issuance, Redemption Rates and the Insolvency Boundary of Deferred Revenue},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论