5D Library › 订阅制单位经济学、净收入留存与人工智能推理标度律 · English

第一章 客户终身价值不是毛利除以流失率

口号及其隐含假设

每一份订阅制路演都有同一行字:终身价值等于每位客户的月毛利除以月流失率。流失率 5% 时,一位客户「值」二十个月的毛利。这个公式在一个假设下是精确的:每一位客户每个月的流失概率都相同。那时流失时间服从几何分布,均值为 1/c1/c,未折现的价值为 GP/c\text{GP}/c。

真实的客户各不相同。有些人从来就不会留下;有些人会留好几年。当流失倾向因人而异时,一个队列逐月的留存率会随时间上升,因为最可能离开的人先走了。人口学家称之为脆弱性选择 (Vaupel et al. 1979);在营销学中,这正是 Fader 与 Hardie 的移位贝塔-几何模型背后的观察:它让每位客户的流失概率 θ\theta 服从一个贝塔分布 (Fader and Hardie 2007);更早的客户计数模型也以同一思路为基础 (Schmittlein et al. 1987)。用客户未来毛利的折现值而不是单一流失率来给企业估值,至少可以追溯到 Gupta、Lehmann 与 Stuart (Gupta et al. 2004)。

推导

移位贝塔-几何模型。 若 θ∼Beta(a,b)\theta \sim \text{Beta}(a, b),客户存活 tt 个月的概率为 S(t)=B(a,b+t)B(a,b),S(t) = \frac{B(a, b + t)}{B(a, b)}, a>1a > 1 时期望寿命为 (a+b−1)/(a−1)(a + b – 1)/(a – 1)——a≤1a \le 1 时则为无穷大。以月利率 dd 折现,队列成员的价值为 ∑tGPS(t)/(1+d)t\sum_t \text{GP}\,S(t)/(1+d)^t,它总是有限的。

推导

# Chapter 1. Why "LTV = margin x ARPU / churn" misleads when customers differ.
# Each customer has a fixed monthly churn probability theta; across customers
# theta follows a Beta(a, b) distribution (the shifted-beta-geometric model of
# Fader and Hardie). A cohort's survival after t months is then S(t) = B(a, b
# + t) / B(a, b), and its month-to-month retention RISES with age because the
# likeliest churners leave first. Monthly gross profit per customer 100 (ARPU
# x margin); monthly discount rate 1 %. Three "naive" LTVs are compared with
# the true discounted cohort LTV:
#   month-1 churn     the churn observed in the first month, used for ever
#   blended churn     the churn of the whole current customer base of a
#   business that
#                     has acquired the same number of customers every month
#                     for 36 or for 120 months (a young and a mature base)
#   pitch-deck LTV    the undiscounted 100 / c of the slogan, using the mature
#   base's churn
import math

GP, D = 100.0, 0.01

def lbeta(a, b):
    return math.lgamma(a) + math.lgamma(b) - math.lgamma(a + b)

def surv(a, b, t):
    return math.exp(lbeta(a, b + t) - lbeta(a, b))

def true_ltv(a, b, horizon=2400):
    return sum(GP * surv(a, b, t) / (1 + D) ** t for t in range(horizon))

def geo_ltv(c):
    return GP * (1 + D) / (c + D)

def blended_churn(a, b, months=36):
    alive = sum(surv(a, b, t) for t in range(months))
    leaving = sum(surv(a, b, t) - surv(a, b, t + 1) for t in range(months))
    return leaving / alive

if __name__ == "__main__":
    print("   Beta(a, b)     churn: month 1   base 36 mo   base 120 mo")
    rows = ((0.7, 8.0), (1.2, 15.0), (2.0, 30.0), (20.0, 300.0))
    for a, b in rows:
        print(
            "   ({:4.1f}, {:5.1f})          {:6.3f}       {:6.3f}      "
            " {:6.3f}".format(
                a,
                b,
                a / (a + b),
                blended_churn(a, b, 36),
                blended_churn(a, b, 120),
            )
        )
    print(
        "\n   LTV (discounted), from a geometric model fitted to the churn"
        " rate"
    )
    print(
        "   of month 1, of the base at 36 months and of the base at 120"
        " months;"
    )
    print("   pitch deck = 100 / c(120 mo), undiscounted")
    print("   (a, b)          true   month-1       base 36 mo    base 120 mo")
    for a, b in rows:
        tl = true_ltv(a, b)
        c1, c36, c120 = (
            a / (a + b),
            blended_churn(a, b, 36),
            blended_churn(a, b, 120),
        )
        print(
            "   ({:4.1f}, {:5.1f})   {:5.0f}   {:5.0f} ({:+4.0%})  {:5.0f}"
            " ({:+4.0%})  {:5.0f} ({:+4.0%})".format(
                a,
                b,
                tl,
                geo_ltv(c1),
                geo_ltv(c1) / tl - 1,
                geo_ltv(c36),
                geo_ltv(c36) / tl - 1,
                geo_ltv(c120),
                geo_ltv(c120) / tl - 1,
            )
        )
    print("\n   (a, b)          true   pitch deck")
    for a, b in rows:
        tl = true_ltv(a, b)
        c120 = blended_churn(a, b, 120)
        print("   ({:4.1f}, {:5.1f})   {:5.0f}   {:5.0f} ({:+4.0%})".format(
            a, b, tl, GP / c120, GP / c120 / tl - 1))
    print(
        "\n   expected lifetime in months, undiscounted: (a + b - 1) / (a"
        " - 1)"
    )
    print("   when a > 1")
    for a, b in ((0.7, 8.0), (1.2, 15.0), (2.0, 30.0), (20.0, 300.0)):
        print(
            "   ({:4.1f}, {:5.1f}): {}".format(
                a,
                b,
                (
                    "infinite (a <= 1)"
                    if a <= 1
                    else "{:.1f}".format((a + b - 1) / (a - 1))
                ),
            )
        )
    print(
        "\nReading: with heterogeneous customers no single churn number gives"
        " the right"
    )
    print(
        "lifetime value. Early churn understates it; the churn of an old"
        " base can"
    )
    print(
        "overstate it; the error changes sign with the age of the base used."
        " Dropping"
    )
    print(
        "the discount on top of a mature base's churn - the usual pitch-deck"
        " number -"
    )
    print("overstates it in every row.")
#    Beta(a, b)     churn: month 1   base 36 mo   base 120 mo
#    ( 0.7,   8.0)           0.080        0.039        0.025
#    ( 1.2,  15.0)           0.074        0.046        0.034
#    ( 2.0,  30.0)           0.062        0.047        0.039
#    (20.0, 300.0)           0.062        0.060        0.060
#
#    LTV (discounted), from a geometric model fitted to the churn rate
#    of month 1, of the base at 36 months and of the base at 120 months;
#    pitch deck = 100 / c(120 mo), undiscounted
#    (a, b)          true   month-1       base 36 mo    base 120 mo
#    ( 0.7,   8.0)    2728    1117 (-59%)   2069 (-24%)   2909 ( +7%)
#    ( 1.2,  15.0)    2152    1201 (-44%)   1805 (-16%)   2293 ( +7%)
#    ( 2.0,  30.0)    1975    1393 (-29%)   1781 (-10%)   2075 ( +5%)
#    (20.0, 300.0)    1443    1393 ( -3%)   1434 ( -1%)   1451 ( +1%)
#
#    (a, b)          true   pitch deck
#    ( 0.7,   8.0)    2728    4045 (+48%)
#    ( 1.2,  15.0)    2152    2938 (+37%)
#    ( 2.0,  30.0)    1975    2586 (+31%)
#    (20.0, 300.0)    1443    1678 (+16%)
#
#    expected lifetime in months, undiscounted: (a + b - 1) / (a - 1)
#    when a > 1
#    ( 0.7,   8.0): infinite (a <= 1)
#    ( 1.2,  15.0): 76.0
#    ( 2.0,  30.0): 31.0
#    (20.0, 300.0): 16.8
#
# Reading: with heterogeneous customers no single churn number gives the right
# lifetime value. Early churn understates it; the churn of an old base can
# overstate it; the error changes sign with the age of the base used. Dropping
# the discount on top of a mature base's churn - the usual pitch-deck number -
# overstates it in every row.

第一部分说明为什么「流失率」不是一个数。在异质性最强的人群中,首月流失率为 0.080,但一家稳定获客的企业,其整个客户基数的流失率在 36 个月后为 0.039,在 120 个月后为 0.025——基数里满是幸存者。

第二部分说明其后果。用首月流失率,会把真实的折现终身价值在异质性最强的情形下低估 −59%;用年轻基数的流失率,低估 −24%;用成熟基数的流失率,则略为高估(+7%)。误差的符号随所用基数的年龄而改变。而通常的路演数字——不折现、用成熟基数的流失率——在每一行都高估了价值,幅度从 +48% 到 +16%。只有近乎同质的最后一行,无论用哪种流失率都接近真值。

寿命那几行是算术上的意外:a≤1a \le 1 时期望寿命为无穷大。拥有这类客户的企业没有一个有限的「寿命」可报,它的价值完全由折现率决定。此时未折现的 LTV 不只是不准确,而是没有定义。

用头六个月作预测

实际的问题是:对一个年轻的队列该怎么办。第二段程序从异质人群中抽取 1,000 位客户,观察六个月,并用两个模型拟合所看到的数据。

推导

# Chapter 1. Projecting a cohort from its first six months. 1,000 customers
# are drawn from a heterogeneous population (each with its own monthly churn
# theta ~ Beta(0.7, 8), as the first row of p01) and watched for 6 months. Two
# models are fitted to those six months by maximum likelihood (grid search)
# and asked to project survival at 12, 24 and 60 months and the discounted
# gross profit of the cohort over 60 months (100 a month per survivor, 1 % a
# month discount):
#   geometric        one churn rate for everybody
#   shifted-beta-geometric (sBG)  churn rates spread as Beta(a, b)
import math
import random

def lbeta(a, b):
    return math.lgamma(a) + math.lgamma(b) - math.lgamma(a + b)

def sbg_s(a, b, t):
    return math.exp(lbeta(a, b + t) - lbeta(a, b))

def simulate(n, a, b, months, rng):
    # lost[t] = churned at end of month t (t = 1..months)
    lost = [0] * (months + 1)
    for _ in range(n):
        th = rng.betavariate(a, b)
        for t in range(1, months + 1):
            if rng.random() < th:
                lost[t] += 1
                break
    alive = n - sum(lost)
    return lost, alive

def ll_sbg(a, b, lost, alive, months):
    ll = 0.0
    for t in range(1, months + 1):
        p = sbg_s(a, b, t - 1) - sbg_s(a, b, t)
        ll += lost[t] * math.log(p)
    return ll + alive * math.log(sbg_s(a, b, months))

def value(surv_fn, months=60):
    return sum(100 * surv_fn(t) / 1.01 ** t for t in range(months))

if __name__ == "__main__":
    rng = random.Random(58002)
    A, B, M = 0.7, 8.0, 6
    lost, alive = simulate(1000, A, B, M, rng)
    n_exposure = sum((1000 - sum(lost[:t])) for t in range(1, M + 1))
    c_geo = sum(lost) / n_exposure
    best = None
    for i in range(1, 80):
        for j in range(1, 160):
            a, b = 0.05 * i, 0.25 * j
            ll = ll_sbg(a, b, lost, alive, M)
            if best is None or ll > best[0]:
                best = (ll, a, b)
    _, ah, bh = best
    print(
        "observed: {} of 1,000 still active after {} months".format(alive, M)
    )
    print(
        "fitted geometric churn {:.4f}; fitted sBG a = {:.2f}, b = {:.2f}"
        .format(c_geo, ah, bh)
    )
    print(
        "\n   survival at month       12       24       60    60-month value"
    )
    print("                                                    per customer")
    for name, fn in (("truth", lambda t: sbg_s(A, B, t)),
                     ("sBG fit", lambda t: sbg_s(ah, bh, t)),
                     ("geometric fit", lambda t: (1 - c_geo) ** t)):
        print(
            "   {:<15}       {:6.3f}   {:6.3f}   {:6.3f}        {:7.0f}"
            .format(name, fn(12), fn(24), fn(60), value(fn))
        )
# observed: 681 of 1,000 still active after 6 months
# fitted geometric churn 0.0632; fitted sBG a = 0.70, b = 8.25
#
#    survival at month       12       24       60    60-month value
#                                                     per customer
#    truth                  0.522    0.375    0.221           1966
#    sBG fit                0.529    0.381    0.225           1991
#    geometric fit          0.457    0.209    0.020           1364

两个模型都能拟合六个月的数据。它们的预测却立刻分道扬镳:到第 60 个月,单一流失率为 0.0632 的几何模型预测队列剩下 0.020;贝塔模型预测 0.225;真值为 0.221。几何模型给出每位客户 1364 的价值,真值为 1966;贝塔模型给出 1991。

队列 60 个月的存活:真值、用头六个月(阴影)拟合的贝塔模型,以及用同样几个月拟合的单一流失率。

当异质性模型也错了的时候

这个结果对贝塔模型有利,因为数据就是由它生成的。第三段程序改变了真相。

推导

# Chapter 1. The same six-month projection when the truth is NOT a beta
# mixture. Two alternative truths, each with about the same six-month
# retention as p02:
#   two segments   25 % loyal customers (monthly churn 0.01) and 75 % fickle
#   (0.11) Weibull        discrete Weibull survival S(t) = exp(-(t /
#   18)^0.55): every customer's
#                  hazard falls with tenure (habit formation), not only the
#                  mix
# 1,000 customers each, observed for 6 months, then the two models of p02 are
# fitted and projected. Value = discounted gross profit over 60 months, 100 a
# month, 1 %/month.
import math
import random

def lbeta(a, b):
    return math.lgamma(a) + math.lgamma(b) - math.lgamma(a + b)

def sbg_s(a, b, t):
    return math.exp(lbeta(a, b + t) - lbeta(a, b))

def weib(t):
    return math.exp(-(t / 18.0) ** 0.55)

def seg(t):
    return 0.25 * 0.99 ** t + 0.75 * 0.89 ** t

def draw(surv, n, months, rng):
    lost = [0] * (months + 1)
    for _ in range(n):
        u = rng.random()
        for t in range(1, months + 1):
            if u > surv(t):
                lost[t] += 1
                break
    return lost, n - sum(lost)

def fit(lost, alive, M):
    expo = sum(1000 - sum(lost[:t]) for t in range(1, M + 1))
    c = sum(lost) / expo
    best = None
    for i in range(1, 80):
        for j in range(1, 160):
            a, b = 0.05 * i, 0.25 * j
            ll = sum(
                lost[t] * math.log(sbg_s(a, b, t - 1) - sbg_s(a, b, t))
                for t in range(1, M + 1)
            )
            ll += alive * math.log(sbg_s(a, b, M))
            if best is None or ll > best[0]:
                best = (ll, a, b)
    return c, best[1], best[2]

def value(fn):
    return sum(100 * fn(t) / 1.01 ** t for t in range(60))

if __name__ == "__main__":
    rng = random.Random(58003)
    for name, truth in (("two segments", seg), ("Weibull k = 0.55", weib)):
        lost, alive = draw(truth, 1000, 6, rng)
        c, a, b = fit(lost, alive, 6)
        print("{}: {} of 1,000 active after 6 months;".format(name, alive))
        print("   sBG fit a = {:.2f}, b = {:.2f}".format(a, b))
        print(
            "   survival at month       12       24       60    60-month"
            " value"
        )
        for lab, fn in (
            ("truth", truth),
            ("sBG fit", lambda t: sbg_s(a, b, t)),
            ("geometric fit", lambda t: (1 - c) ** t),
        ):
            v = value(fn)
            print(
                "   {:<15}       {:6.3f}   {:6.3f}   {:6.3f}     {:6.0f}"
                " ({:+4.0%})".format(
                    lab, fn(12), fn(24), fn(60), v, v / value(truth) - 1
                )
            )
        print()
    print("Reading: one churn rate understates in both"
          " worlds. The heterogeneity")
    print("model is close when its assumption roughly"
          " holds and overstates when the")
    print("hazard itself falls with tenure. Six months"
          " cannot tell these worlds")
    print("apart, so a 60-month value should be"
          " reported as a range across models,")
    print("not as one number.")
# two segments: 605 of 1,000 active after 6 months;
#    sBG fit a = 1.40, b = 13.75
#    survival at month       12       24       60    60-month value
#    truth                  0.407    0.242    0.137       1513 ( +0%)
#    sBG fit                0.419    0.246    0.097       1472 ( -3%)
#    geometric fit          0.360    0.130    0.006       1099 (-27%)
#
# Weibull k = 0.55: 583 of 1,000 active after 6 months;
#    sBG fit a = 0.25, b = 1.25
#    survival at month       12       24       60    60-month value
#    truth                  0.449    0.310    0.144       1631 ( +0%)
#    sBG fit                0.516    0.438    0.350       2170 (+33%)
#    geometric fit          0.310    0.096    0.003        979 (-40%)
#
# Reading: one churn rate understates in both worlds. The heterogeneity
# model is close when its assumption roughly holds and overstates when the
# hazard itself falls with tenure. Six months cannot tell these worlds
# apart, so a 60-month value should be reported as a range across models,
# not as one number.

在两个细分群体的世界里——四分之一忠诚,四分之三善变——贝塔模型仍然接近(−3%),几何模型则不然(−27%)。在韦布尔真相下,每位客户的风险率都随使用时长而下降(是习惯养成,而不是选择),贝塔模型把 60 个月的价值高估 +33%,几何模型低估 −40%。六个月的数据无法区分这两个世界。

注

应当报告什么。 单一的终身价值,是一个模型给出的点估计,而这个模型的结构在早期无法用数据检验。诚实的报告是:多个合理模型给出的区间、所用的折现率,以及数据的时长。一个不说明由哪个模型得出的预测,不是一个数字,而是一个假设。

假说 · 尚未验证

作者的假说(尚未验证): 在报告队列留存的订阅制企业中,绝大多数队列的逐月留存率随队列年龄上升;以首季流失率计算的 LTV 低估实际的五年队列价值,而以混合基数流失率计算的 LTV 对新渠道获取的队列则会高估。什么结果会证明它错:多数队列的逐月留存率持平,或者误差没有系统性的符号。

小结

「LTV = 毛利 / 流失率」假设每位客户以相同的速率流失。客户异质时,流失率随队列年龄下降,任何单一流失率 LTV 的误差符号随所取基数的年龄而改变,而常见的路演数字——不折现、用成熟基数的流失率——在例子中把价值高估 16% 到 48%;当贝塔参数 a≤1a \le 1 时,期望寿命为无穷大,只有折现才能给出有限的价值。结构正确时,异质性模型的预测远好于单一流失率;结构不对时,它会高估三分之一;六个月的数据无法分辨,所以终身价值应当报告为跨模型的区间。

参考文献

Fader, Peter S., and Bruce G. S. Hardie. 2007. “How to Project Customer Retention.” Journal of Interactive Marketing 21 (1): 76–90.
Gupta, Sunil, Donald R. Lehmann, and Jennifer Ames Stuart. 2004. “Valuing Customers.” Journal of Marketing Research 41 (1): 7–18.
Schmittlein, David C., Donald G. Morrison, and Richard Colombo. 1987. “Counting Your Customers: Who Are They and What Will They Do Next?” Management Science 33 (1): 1–24.
Vaupel, James W., Kenneth G. Manton, and Eric Stallard. 1979. “The Impact of Heterogeneity in Individual Frailty on the Dynamics of Mortality.” Demography 16 (3): 439–54.

Fan, Y. (2026). Subscription Unit Economics, Net Revenue Retention and AI Inference Scaling Laws: Cohort Survival Mixtures, Negative-Churn Differential Equations, Token Compute Margins and Metcalfe Congestion Limits [订阅制单位经济学、净收入留存与人工智能推理标度律] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090164 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).