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

第八章 完整账本、区分棘轮与筛选的实验,以及预注册的留出实验

挽回流程的完整账本

推导

# Chapter 8. p14: The full ledger of a three-message recovery flow per
# 1,000 abandoners (schedule 0.25 h / 24 h / 72 h, model of p01).
# Incremental orders only; recovered orders are refunded at rate 0.06;
# unsubscribes as in p02; a spam complaint (0.001 per message) costs
# RM 200 in lost deliverability. Compared with what attribution claims.
import importlib.util

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

p01, p02, p04 = load("p01_return_hazard"), load("p02_how_many"), \
    load("p04_incremental")
RULE, N, ORDER = (0.25, 24, 72), 1000, 98.0
REFUND, SPAM, SPAM_COST = 0.06, 0.001, 200.0

if __name__ == "__main__":
    claimed = p04.after_first(RULE) * ORDER
    inc = N * (p01.returned(RULE) - p01.returned(())) * ORDER
    refunds = inc * REFUND
    unsub = N * (1 - (1 - p02.U) ** 3) * p02.V
    spam = N * 3 * SPAM * SPAM_COST
    net = inc - refunds - unsub - spam
    print("   attribution claims  RM {:8,.0f}".format(claimed))
    print("   incremental revenue RM {:8,.0f}".format(inc))
    print("   refunds            -RM {:8,.0f}".format(refunds))
    print("   unsubscribes       -RM {:8,.0f}".format(unsub))
    print("   spam complaints    -RM {:8,.0f}".format(spam))
    print("   net                 RM {:8,.0f} ({:.0%} of the claim)".format(
        net, net / claimed))
#    attribution claims  RM   35,672
#    incremental revenue RM   15,207
#    refunds            -RM      912
#    unsubscribes       -RM      717
#    spam complaints    -RM      600
#    net                 RM   12,978 (36% of the claim)

按 0.25 小时、24 小时、72 小时三封跟进,每 1,000 名弃单者:归因数字声称 RM 35,672;跟进真正带来的增量收入 RM 15,207;扣除挽回订单的退款 RM 912、退订造成的未来价值损失 RM 717 与垃圾邮件投诉的送达损失 RM 600,净收入 RM 12,978,只有归因数字的 36%。这仍然是一笔正收入,但只有声称的三分之一多一点。

区分棘轮与筛选

推导

# Chapter 8. p15: Telling a ratchet from selection. Cohort curves cannot
# (p06). An experiment can: randomly give half of new subscribers a free
# import of 12 months of history. Under the ratchet of p05 their first
# renewal rises from 0.801 to 0.958; under selection it stays at 0.806.
# Sample size per arm (two-sided 0.05, power 0.80) to detect a smaller
# real effect, if only part of the ratchet is real.
from math import ceil, sqrt
from statistics import NormalDist

N = NormalDist()

def n_per_arm(p1, p2, alpha=0.05, power=0.80):
    za, zb = N.inv_cdf(1 - alpha / 2), N.inv_cdf(power)
    pbar = (p1 + p2) / 2
    num = (za * sqrt(2 * pbar * (1 - pbar))
           + zb * sqrt(p1 * (1 - p1) + p2 * (1 - p2))) ** 2
    return ceil(num / (p1 - p2) ** 2)

if __name__ == "__main__":
    for p2 in (0.958, 0.85, 0.82):
        print("   renewal 0.801 vs {:.3f}: {:,} subscribers per arm".format(
            p2, n_per_arm(0.801, p2)))
#    renewal 0.801 vs 0.958: 67 subscribers per arm
#    renewal 0.801 vs 0.850: 941 subscribers per arm
#    renewal 0.801 vs 0.820: 6,678 subscribers per arm

第三章说明群组曲线无法区分棘轮与筛选,但实验可以:随机给一半新订阅者免费导入 12 个月的历史记录,比较第一次续费率。若棘轮完全为真(0.801 对 0.958),每组只要 67 人;若只有一部分为真(0.801 对 0.850),每组 941 人;若效应很小(0.801 对 0.820),每组 6,678 人。

注

四步核查。 第一,弃单跟进只发给同意接收营销信息的买家,并在每封中提供退订 (Parliament of Malaysia 2010)。第二,折扣券只给首次购买者,对回头客发普通提醒。第三,续费策略以第八章的随机实验为准,不以群组曲线为准。第四,挽回、推荐与顾问业绩一律按留出组度量增量,并以退款率为护栏 (Kohavi et al. 2020)。

预注册的留出实验

推导

# Chapter 8. p16: A pre-registered holdout for the recovery flow.
# Abandoners are randomised: most get the three messages, a holdout gets
# none. Primary outcome: an order within 7 days. Detect 0.214 vs 0.30
# (about half of the modelled gain of 0.155), two-sided 0.05, power 0.80,
# with unequal arms.
from math import ceil
from statistics import NormalDist

N = NormalDist()

def n_holdout(p1, p2, ratio, alpha=0.05, power=0.80):
    za, zb = N.inv_cdf(1 - alpha / 2), N.inv_cdf(power)
    var = p1 * (1 - p1) / ratio + p2 * (1 - p2)
    return ceil((za + zb) ** 2 * var / (p1 - p2) ** 2)

if __name__ == "__main__":
    print("   outcome: order within 7 days; 0.30 vs 0.214")
    for share in (0.5, 0.2, 0.1):
        ratio = (1 - share) / share
        n2 = n_holdout(0.30, 0.214, ratio)
        print("   holdout {:.0%}: holdout {:,} abandoners, total {:,}".format(
            share, n2, ceil(n2 * (1 + ratio))))
#    outcome: order within 7 days; 0.30 vs 0.214
#    holdout 50%: holdout 402 abandoners, total 804
#    holdout 20%: holdout 235 abandoners, total 1,175
#    holdout 10%: holdout 204 abandoners, total 2,040

弃单者被随机分成两组:大部分收到三封跟进,留出组不收任何跟进。主要结局为 7 天内下单,检验 0.30 对 0.214。留出组占一半时需留出组 402 人、共 804 人;占两成时留出组 235 人、共 1,175 人;占一成时留出组 204 人、共 2,040 人。

假说 · 尚未验证

预注册的预测(尚未验证): 三封跟进的增量回访率不超过归因回访率的一半;随机导入历史记录不会使第一次续费率提高 5 个百分点以上。什么会证明它错: 增量回访率显著超过归因的一半,或导入组的第一次续费率显著高出 5 个百分点以上。

全书小结

弃单跟进的最优时序很平,要紧的是第一封尽早发出;跟进的边际收益迅速递减。折扣钩子会教会回头客等待;归因挽回大大高估了效果。上升的续费曲线既可能来自数据棘轮,也可能完全来自筛选,只有随机实验能够区分;在此之前,不应以「锁定」为由拒绝数据导出。高客单产品的甄别靠菜单的价格差、对两类客户难易不同的问题与有限顾问时间的排序,而不是泛泛的摩擦或假紧迫感。

参考文献

Kohavi, Ron, Diane Tang, and Ya Xu. 2020. Trustworthy Online Controlled Experiments: A Practical Guide to a/b Testing. Cambridge University Press.
Parliament of Malaysia. 2010. Personal Data Protection Act 2010 (Act 709). Laws of Malaysia.

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).