5D Library › 回溯信念锁定、序贯痛点唤醒与马尔可夫升单自动机 · English
第三章 承诺棘轮:价格跳升比与升单时机
两条已发表的规律
升单(加购、一键升级、课程转化)依赖两条有实验证据的规律。其一是「登门槛」:先答应一个小请求的人,更可能答应随后的大请求 (Freedman and Fraser 1966)。其二是沉没成本:已经投入的人更倾向于继续投入,即使从当下看继续并不划算 (Arkes and Blumer 1985);在组织决策中,这表现为对既定方向的承诺升级 (Staw 1976)。
证据
两条规律的出处。 Freedman 与 Fraser 的实验中,先同意在窗口贴一张小标语的住户,更多地同意随后在院子里立一块大牌子 (Freedman and Fraser 1966)。Arkes 与 Blumer 的实验显示,已经付了更多钱的人更可能继续使用他们事后后悔购买的东西 (Arkes and Blumer 1985)。Staw 显示,对一项失败决策负有个人责任的人,会向它追加更多资源 (Staw 1976)。这些实验说明效应存在,但没有给出可以直接用于定价的参数;本章的参数都是示意值。
一次升单的最优跳升比
设用户已购第 阶(价格 ),对第 阶(价格 )的接受概率为 其中 是抗拒阈值,随此前的累计花费 (沉没成本)上升。接受者中有一部分会要求退款,被推得越狠,比例越高。
推导
# Chapter 3. [p05] The optimal price jump for one upsell, with and
# without the cost of regret. A buyer of tier i (price p) is offered
# tier i+1 at r * p. She accepts with logistic probability
# A(r) = 1 / (1 + exp((r - r0) / 0.6)),
# where her resistance threshold r0 = 2.5 + 0.8 * ln(1 + S / p) rises
# with what she has already spent, S (sunk cost; S = 0 for a first
# purchase). A share of buyers who accept ask for a refund, more so
# the harder they were pushed: phi(r) = 0.05 + 0.30 / (1 + exp(-(r -
# r0) / 0.6)). Net revenue per offer, in units of p: r A(r) (1 - phi).
import math
def r0(sunk):
return 2.5 + 0.8 * math.log(1 + sunk)
def accept(r, t):
return 1 / (1 + math.exp((r - t) / 0.6))
def refund(r, t):
return 0.05 + 0.30 / (1 + math.exp(-(r - t) / 0.6))
def best(sunk, with_refunds):
t = r0(sunk)
grid = [1 + k / 100 for k in range(0, 1201)]
def net(r):
return r * accept(r, t) * ((1 - refund(r, t)) if with_refunds
else 1.0)
r = max(grid, key=net)
return t, r, accept(r, t), refund(r, t), net(r)
if __name__ == "__main__":
print(" sunk S/p threshold no refunds: r* A(r*) with refunds:"
" r* A(r*) phi")
for sunk in (0, 1, 3, 10):
t, r1, a1, f1, n1 = best(sunk, False)
t, r2, a2, f2, n2 = best(sunk, True)
print(" {:7d} {:10.3f} {:16.2f} {:6.3f} {:17.2f} {:6.3f} {:5.3f}"
.format(sunk, t, r1, a1, r2, a2, f2))
print("\nReading: sunk cost raises the jump a buyer tolerates, so the")
print("best ratio grows along the ladder. The optimum always lies")
print("well below the resistance threshold, and counting refunds")
print("lowers it further: pushing to the threshold does not pay.")
# sunk S/p threshold no refunds: r* A(r*) with refunds: r* A(r*) phi
# 0 2.500 1.99 0.701 1.86 0.744 0.127
# 1 3.055 2.40 0.749 2.26 0.790 0.113
# 3 3.609 2.82 0.788 2.68 0.825 0.103
# 10 4.418 3.48 0.827 3.33 0.860 0.092
#
# Reading: sunk cost raises the jump a buyer tolerates, so the
# best ratio grows along the ladder. The optimum always lies
# well below the resistance threshold, and counting refunds
# lowers it further: pushing to the threshold does not pay.第一次购买()时,抗拒阈值为 2.500,最优跳升比为 1.99,计入退款后为 1.86;累计花费为当前价格的 10 倍时,阈值升到 4.418,最优跳升比为 3.48(计入退款后 3.33)。最优点总在阈值之下很远的地方:在阈值处,一半的报价被拒,接受者中后悔的比例也最高。所以「把价格推到用户能承受的极限」并不是利润最大化,计入退款后更不是。
四级阶梯:比例与时机
再看一条四级阶梯:第一阶 9,之后每次升单在上次购买之后 天发出。用户的「准备度」在感受到第一次结果时达到高峰(这里设为第 7 天),同时用户每天都有一定的概率离开。程序同时搜索三个跳升比与共同的时机。
推导
# Chapter 3. [p06] Timing and ratios along a four-step ladder.
# Tier 1 costs 9. After each purchase the next offer is made t days
# later. Readiness peaks when the first result has been felt:
# g(t) = (t/7) exp(1 - t/7); the buyer is still around with probability
# exp(-t/60). Acceptance of the next step is g(t) * A(r) with A and the
# refund share phi from p05 (sunk cost counted). The ratios r1, r2, r3
# and one common delay t are chosen to maximise expected net revenue
# from the three upsells (grid search).
import math
from itertools import product
def thr(sunk):
return 2.5 + 0.8 * math.log(1 + sunk)
def acc(r, t):
return 1 / (1 + math.exp((r - t) / 0.6))
def phi(r, t):
return 0.05 + 0.30 / (1 + math.exp(-(r - t) / 0.6))
def ready(t):
return (t / 7) * math.exp(1 - t / 7) * math.exp(-t / 60)
def value(rs, t):
p, spent, alive, total = 9.0, 9.0, 1.0, 0.0
for r in rs:
th = thr(spent / p - 1)
a = ready(t) * acc(r, th)
alive *= a
price = r * p
total += alive * price * (1 - phi(r, th))
spent += price
p = price
return total
if __name__ == "__main__":
grid = [1.5 + k / 10 for k in range(0, 31)]
best = max(((value(rs, t), rs, t) for rs in product(grid, repeat=3)
for t in (1, 3, 5, 7, 10, 14, 21, 30)),
key=lambda x: x[0])
v, rs, t = best
prices = [9.0]
for r in rs:
prices.append(prices[-1] * r)
print(" best delay t = {} days; ratios r1..r3 = {}".format(
t, ", ".join("{:.1f}".format(r) for r in rs)))
print(" prices: " + " -> ".join("{:.2f}".format(x) for x in prices))
print(" expected net upsell revenue per tier-1 buyer: {:.2f}".format(v))
for tt in (1, 7, 30):
print(" same ratios, delay {:2d} days: {:.2f}".format(
tt, value(rs, tt)))
print("\nReading: the ratios rise along the ladder as sunk cost")
print("accumulates, and timing matters as much as price: an offer")
print("the next day or a month later earns far less than one made")
print("when the first result has been felt.")
# best delay t = 7 days; ratios r1..r3 = 1.9, 2.1, 2.2
# prices: 9.00 -> 17.10 -> 35.91 -> 79.00
# expected net upsell revenue per tier-1 buyer: 45.35
# same ratios, delay 1 days: 6.67
# same ratios, delay 7 days: 45.35
# same ratios, delay 30 days: 1.25
#
# Reading: the ratios rise along the ladder as sunk cost
# accumulates, and timing matters as much as price: an offer
# the next day or a month later earns far less than one made
# when the first result has been felt.最优时机是第 7 天,跳升比依次为 1.9、2.1、2.2,价格从 9.00 升到 79.00,每位第一阶买家带来 45.35 的升单净收入。同样的比例,若第二天就发出,只有 6.67;若一个月后才发出,只有 1.25。时机与价格同样重要:太早,用户还没有感受到第一次的结果;太晚,用户已经离开。
假说 · 尚未验证
作者的诠释(尚未验证): 升单的利润主要来自时机(在用户感受到结果之后提出),而不是来自把价格推向用户的承受极限。什么会证明这一判断错:若在随机分配发出时机的实验中,第 1 天与第 7 天发出的升单接受率没有显著差别,说明「准备度」高峰并不存在,本章的时机结论不成立。
小结
沉没成本使用户能承受的跳升比沿阶梯上升,但最优跳升比始终远低于抗拒阈值,计入退款后更低。在四级阶梯中,最优时机在用户感受到结果之时;同样的价格,早发或晚发都只能拿到一小部分收入。
参考文献
Fan, Y. (2026). Retrodictive Belief Locking, Sequential Pain-Point Agitation and Markovian Upsell Automata: Bayesian Trust Avalanches, Sunk-Cost Commitment Ratchets and Cross-Domain Transition Matrices [回溯信念锁定、序贯痛点唤醒与马尔可夫升单自动机] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090186 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).