5D Library › 弃单挽回动力学、沉没数据续费棘轮与高客单顾问筛选机制 · English
第三章 沉没数据的续费棘轮,与同样上升的筛选曲线
棘轮模型
订阅者在测算软件里保存的历史记录、笔记与图表越多,换到别处的代价就越高。把这种转换成本写成 (以 RM 计, 为月),每月续费概率为 。这里所有订阅者都相同,续费率上升完全来自每个人自己积累的数据。转换成本如何改变竞争与定价,是产业组织理论的经典题目 (Klemperer 1987)。
推导
# Chapter 3. p05: A sunk-data ratchet. A subscriber's saved history
# (readings, notes, charts) has a switching cost that accumulates,
# s(t) = S (1 - exp(-t / T)) in RM, t in months. Each month the
# subscriber renews with probability logistic(c + s(t) / w - x), where x
# is the price relative to w. All subscribers are alike (no selection).
from math import exp
S, T, W = 120.0, 6.0, 60.0
C, X = 1.39, 0.0 # renewal 0.80 at t = 0 without history
def s(t):
return S * (1 - exp(-t / T))
def renew(t, ratchet=True):
z = C + (s(t) / W if ratchet else 0.0) - X
return 1 / (1 + exp(-z))
def survival(months, ratchet=True):
alive, out = 1.0, []
for t in range(months):
alive *= renew(t, ratchet)
out.append(alive)
return out
def tenure(ratchet=True, months=240):
return 1 + sum(survival(months, ratchet))
if __name__ == "__main__":
for t in (0, 3, 6, 12, 24):
print(" month {:2d}: switching cost RM {:5.1f},"
" renewal {:.3f}".format(t, s(t), renew(t)))
print(" expected months paid: with ratchet {:.1f},"
" without {:.1f}".format(tenure(True), tenure(False)))
# month 0: switching cost RM 0.0, renewal 0.801
# month 3: switching cost RM 47.2, renewal 0.898
# month 6: switching cost RM 75.9, renewal 0.934
# month 12: switching cost RM 103.8, renewal 0.958
# month 24: switching cost RM 117.8, renewal 0.966
# expected months paid: with ratchet 15.9, without 5.0示意参数下,第一个月续费率 0.801;第三个月转换成本 RM 47.2,续费率 0.898;第十二个月 RM 103.8,续费率 0.958;第二十四个月 RM 117.8,续费率 0.966。有棘轮时平均付费 15.9 个月,没有棘轮时只有 5.0 个月。
同一条曲线,另一种解释
推导
# Chapter 3. p06: Rising renewal without any ratchet. In the shifted-
# beta-geometric (sBG) model each subscriber has a fixed monthly churn
# probability theta ~ Beta(a, b); nobody becomes stickier, yet the
# cohort renewal rate (b + t - 1) / (a + b + t - 1) rises because those
# likely to churn leave first. We fit (a, b) to the cohort survival
# curve of the ratchet model in p05 and compare the two curves.
import importlib.util
from math import lgamma, exp
def load(name):
spec = importlib.util.spec_from_file_location(name, name + ".py")
m = importlib.util.module_from_spec(spec)
spec.loader.exec_module(m)
return m
p05 = load("p05_sunk_data")
MONTHS = 24
def lbeta(a, b):
return lgamma(a) + lgamma(b) - lgamma(a + b)
def sbg_survival(a, b, months=MONTHS):
return [exp(lbeta(a, b + t) - lbeta(a, b)) for t in range(1, months + 1)]
def fit(target):
best = None
for i in range(1, 121):
a = i * 0.05
for j in range(1, 241):
b = j * 0.1
sv = sbg_survival(a, b)
err = sum((x - y) ** 2 for x, y in zip(sv, target))
if best is None or err < best[0]:
best = (err, a, b)
return best
if __name__ == "__main__":
target = p05.survival(MONTHS)
err, a, b = fit(target)
sv = sbg_survival(a, b)
gap = max(abs(x - y) for x, y in zip(sv, target))
print(" sBG fit to the ratchet cohort: a = {:.2f}, b = {:.1f}".format(
a, b))
print(" largest survival gap over {} months: {:.3f}".format(
MONTHS, gap))
for t in (1, 4, 7, 13):
rate = (b + t - 1) / (a + b + t - 1)
print(" month {:2d} renewal: ratchet {:.3f}, selection {:.3f}"
.format(t - 1, p05.renew(t - 1), rate))
print(" a new subscriber given 12 months of imported history:")
print(" ratchet model renewal {:.3f}; selection model {:.3f}".format(
p05.renew(12), b / (a + b)))
# sBG fit to the ratchet cohort: a = 0.65, b = 2.7
# largest survival gap over 24 months: 0.011
# month 0 renewal: ratchet 0.801, selection 0.806
# month 3 renewal: ratchet 0.898, selection 0.898
# month 6 renewal: ratchet 0.934, selection 0.930
# month 12 renewal: ratchet 0.958, selection 0.958
# a new subscriber given 12 months of imported history:
# ratchet model renewal 0.958; selection model 0.806移位贝塔几何(sBG)模型假定每个订阅者有一个固定不变的每月流失概率,彼此不同,服从贝塔分布 (Fader and Hardie 2007)。没有任何人变得更黏,可是群组的续费率 仍然逐月上升,因为容易流失的人先走了。用它去拟合上面棘轮模型的群组存活曲线,得到 、,24 个月内两条存活曲线的最大差距只有 0.011;逐月续费率几乎重合(第三个月都是 0.898,第十二个月都是 0.958)。
两种模型对个人的预测却完全不同。给一位新订阅者免费导入 12 个月的历史记录:在棘轮模型中,她的第一次续费率从 0.801 升到 0.958;在筛选模型中,仍然是 0.806,因为数据本来就不是原因。
注
为什么这很要紧。 如果续费上升来自筛选,那么「让用户多存数据」的产品投入不会提高续费,「数据越多越难离开」的说法也只是事后的叙事。群组曲线本身无法区分两种解释;第八章给出一个能区分的随机实验。
假说 · 尚未验证
作者的假说(尚未验证): 作者的订阅产品中,续费率上升主要来自筛选而不是数据积累。什么会证明它错: 随机给一半新订阅者导入历史记录,其第一次续费率显著高于对照组(第八章 p15)。
小结
棘轮模型与筛选模型可以产生几乎相同的群组续费曲线(最大差距 0.011),但对干预的预测截然不同。上升的续费曲线不能证明存在数据棘轮。
参考文献
Fan, Y. (2026). Cart-Abandonment Recovery Kinetics, Sunk-Data Retention Ratchets and High-Ticket B2B Qualification Mechanisms: Multi-Pulse Follow-Up Timing, Switching-Cost Accumulation and Screening Equilibria [弃单挽回动力学、沉没数据续费棘轮与高客单顾问筛选机制] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090209 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).