关于「客户终身毛利:sBG 留存、按批次核算、折现与获客回本期」,《托管月费与价值锚定抽成》给出的定量结论是什么?

结论:终身毛利由三个与单月账单无关的数决定:留多久、钱的时间价值、签下来花了多少。留存要用承认客户之间差异的 sBG 模型去估,常数流失率不是粗一点而是方向错,p50 的两百个批次里它系统性地低估长期留存;七个月的数据已经能给出无偏的估计,但单个批次的误差仍有几个月。留存要按签约批次看,全体流失率主要反映销售节奏,在头两年高估、在十年期限上低估新客户的价值。折现率改变结论三四成,核算期限的影响更大,获客与定价决策用六十个月期限。回本期要按每签一家、折现后计,同时报告收不回来的比例。四个杠杆里,按同样口径比较,提价与降流失最有力,而服务投入只需降低不到一成的流失就能回本;锁定靠的是服务与可携,而不是锁死退出。下一章把这些数字背后的承诺写成合同条款。

数据口径:书中的公司与数字全部虚构,合同与法规只写马来西亚与台湾的一般框架,不是法律、税务或投资意见。 书中以标注框区分已发表的证据、可运行的计算、原典与作者的假说,每条假说都写明什么结果会推翻它。

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

cohort 150 shops; survivors months 0-18:
150 140 130 121 114 107 99 97 93 90 87 83 79 73 70 67 62 61 60
sBG fit on months 1-7: alpha 0.899, beta 11.246, log-likelihood -198.53
mean churn propensity alpha/(alpha+beta)0.0741; geometric fit: churn 0.0616 every month
month seen sBG-fit geometric | retention seen sBG-fit
1140 138.9 140.8 | 0.933 0.926 (fitted)
3121 121.2 124.0 | 0.931 0.936 (fitted)
5107 107.6 109.2 | 0.939 0.944 (fitted)
797 96.9 96.1 | 0.980 0.950 (fitted)
990 88.2 84.7 | 0.968 0.955 (projected)
1279 77.9 70.0 | 0.952 0.961 (projected)
1567 69.8 57.8 | 0.957 0.966 (projected)
1860 63.3 47.8 | 0.984 0.969 (projected)
mean absolute projection error, months 8-18: sBG 2.6 shops, geometric 8.0 shops
expected renewals within 36 months: sBG 16.98, geometric 13.70, true 17.47
expected renewals within 120 months: sBG 30.83, geometric 15.24, true 31.18
share still active at month 36: sBG 0.274, geometric 0.102, true 0.280
200 fresh cohorts: estimated expected renewals within 36 months (true 17.47)
7 months seen, sBG : median17.50, 10th-90th pct 14.65 to 20.64, mean error +0.11
7 months seen, geometric: median14.61, 10th-90th pct 12.75 to 16.60, mean error -2.86
12 months seen, sBG : median17.54, 10th-90th pct 15.88 to 19.65, mean error +0.11
12 months seen, geometric: median15.67, 10th-90th pct 14.36 to 17.33, mean error -1.77
查看完整验算程序
# Chapter 5. p50: Fitting a shifted-beta-geometric (sBG) retention model to one fictional cohort.
# 150 fictional buying-agent shops sign Bridgecalc's monthly hosted plan in the same quarter. Each shop
# has its own monthly churn probability drawn from Beta(1, 14), the planning values of p13; that "truth"
# is invented so the fit can be checked, and because the data are generated by an sBG process the good
# fit proves nothing about real shops. We observe renewals for months 1-7, fit sBG by maximum likelihood
# (right-censored), fit a constant-churn (geometric) model to the same months, and project months 8-18.
# Seed 2007.
# A second block repeats the whole exercise on 200 fresh cohorts (seeds 1-200) with 7 and 12 months seen.
import math
import random

N, ALPHA, BETA, SEEN, LATER = 150, 1.0, 14.0, 7, 18


def cohort(seed):
    rng = random.Random(seed)
    life = []
    for _ in range(N):
        theta = rng.betavariate(ALPHA, BETA)
        t = 1
        while rng.random() >= theta and t <= 60:
            t += 1
        life.append(t)                   # month in which the shop churns (61 = still active at 60)
    return [sum(1 for t in life if t > m) for m in range(LATER + 1)]


def sbg_surv(al, be, horizon):
    s, out = 1.0, [1.0]
    for t in range(1, horizon + 1):
        s *= (be + t - 1) / (al + be + t - 1)
        out.append(s)
    return out


def loglik(al, be, alive, seen):
    s = sbg_surv(al, be, seen)
    terms = [(alive[t - 1] - alive[t]) * math.log(s[t - 1] - s[t]) for t in range(1, seen + 1)]
    return math.fsum(terms) + alive[seen] * math.log(s[seen])


def fit(alive, seen, steps=(0.2, 0.05, 0.01, 0.002), width=15):
    best = None
    for step in steps:                   # coarse-to-fine grid on log(alpha), log(beta)
        c = best[1:] if best else (0.0, 1.0)
        for i in range(-width, width + 1):
            for j in range(-width, width + 1):
                la, lb = c[0] + i * step, c[1] + j * step
                ll = loglik(math.exp(la), math.exp(lb), alive, seen)
                if best is None or ll > best[0]:
                    best = (ll, la, lb)
    exposure = sum(alive[m - 1] for m in range(1, seen + 1))
    return math.exp(best[1]), math.exp(best[2]), best[0], (alive[0] - alive[seen]) / exposure


alive = cohort(2007)
a_hat, b_hat, ll_hat, p_geo = fit(alive, SEEN)
print("cohort {} shops; survivors months 0-{}:".format(N, LATER))
print("   " + " ".join(str(x) for x in alive))
print("sBG fit on months 1-{}: alpha {:.3f}, beta {:.3f}, log-likelihood {:.2f}".format(
    SEEN, a_hat, b_hat, ll_hat))
print("mean churn propensity alpha/(alpha+beta) = {:.4f}; geometric fit: churn {:.4f} every month".format(
    a_hat / (a_hat + b_hat), p_geo))
s_sbg = sbg_surv(a_hat, b_hat, 120)
s_true = sbg_surv(ALPHA, BETA, 120)
print("month  seen  sBG-fit  geometric | retention seen  sBG-fit")
for m in (1, 3, 5, 7, 9, 12, 15, 18):
    ret_seen = alive[m] / alive[m - 1]
    ret_fit = s_sbg[m] / s_sbg[m - 1]
    tag = "  (fitted)" if m <= SEEN else "  (projected)"
    print("{:5d} {:5d} {:8.1f} {:10.1f} | {:14.3f} {:8.3f}{}".format(
        m, alive[m], N * s_sbg[m], N * (1 - p_geo) ** m, ret_seen, ret_fit, tag))
err_s = math.fsum(abs(N * s_sbg[m] - alive[m]) for m in range(SEEN + 1, LATER + 1)) / (LATER - SEEN)
err_g = math.fsum(abs(N * (1 - p_geo) ** m - alive[m]) for m in range(SEEN + 1, LATER + 1)) / (LATER - SEEN)
print("mean absolute projection error, months 8-18: sBG {:.1f} shops, geometric {:.1f} shops".format(
    err_s, err_g))
for H in (36, 120):
    print("expected renewals within {} months: sBG {:.2f}, geometric {:.2f}, true {:.2f}".format(
        H, math.fsum(s_sbg[1:H + 1]), math.fsum((1 - p_geo) ** m for m in range(1, H + 1)),
        math.fsum(s_true[1:H + 1])))
print("share still active at month 36: sBG {:.3f}, geometric {:.3f}, true {:.3f}".format(
    s_sbg[36], (1 - p_geo) ** 36, s_true[36]))
true36 = math.fsum(s_true[1:37])
print("200 fresh cohorts: estimated expected renewals within 36 months (true {:.2f})".format(true36))
for seen in (7, 12):
    est_s, est_g = [], []
    for seed in range(1, 201):
        al = cohort(seed)
        a1, b1, _, pg = fit(al, seen, steps=(0.25, 0.05, 0.01), width=10)
        est_s.append(math.fsum(sbg_surv(a1, b1, 36)[1:]))
        est_g.append(math.fsum((1 - pg) ** m for m in range(1, 37)))
    for name, e in (("sBG", est_s), ("geometric", est_g)):
        e.sort()
        print("   {:2d} months seen, {:9}: median {:6.2f}, 10th-90th pct {:6.2f} to {:6.2f}, "
              "mean error {:+.2f}"
              .format(seen, name, e[100], e[20], e[180], math.fsum(e) / 200 - true36))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《技术底座托管月费、价值锚定抽成混合契约与客户终身毛利锁定母典》第 5 章「客户终身毛利:sBG 留存、按批次核算、折现与获客回本期」(版本 0.1.0). 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-135 · v0.1.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026135,
  author    = {Fan, Yafu},
  title     = {Hosted-Platform Retainers, Value-Anchored Revenue-Share Hybrids and Customer Lifetime Gross-Margin Lock-In},
  year      = {2026},
  publisher = {5D Library},
  edition   = {0.1.0},
  note      = {DOI pending}
}

本书其他章节结论