5D Library › 多阶段最后通牒议价、应收账款衰减动力学与最优施压序列 · English
第八章 实践:催收手册、开价与认识一个客户
小公司的催收手册
p14 模拟一家小公司五十年的逾期发票,金额从 500 到 20,000,比较三种政策。第二种对每张发票使用第四章的时间表;第三种对小额发票(第五章所说威胁会落空的那些)用小额诉讼立案代替律师函,让威胁变得可信。
推导
# Chapter 8. p14: A small firm's collection playbook, simulated. 200
# overdue invoices a year with sizes from 500 to 20,000 (log-uniform),
# debtors and stages as in p07 (stop service day 60, lawyer's letter
# day 105; lost margin 0.3 of the amount for a prompt customer still
# unpaid at either step). The letter costs 5% of the amount. Debtors
# ignore a letter whose threat is empty (amount below 4,762 in an
# ordinary court, p09), so it adds nothing beyond stage 2. Policies:
# 1 reminders only;
# 2 the timetable for every invoice;
# 3 the timetable, but below 4,762 a small-claims filing (cost 300)
# replaces the letter, which makes the threat credible.
# Net recovery as a share of the amount owed over 50 simulated years.
# Seed 94.
import random
from math import exp, log
KINDS = ((0.6, 1 / 15, 0.0, "prompt"), (0.3, 1 / 90, 0.0, "slow"),
(0.1, 1 / 400, 1 / 250, "distressed"))
MULT = {"prompt": (1.5, 3.0, 3.0), "slow": (1.5, 3.0, 6.0),
"distressed": (1.5, 1.5, 3.0)}
NEVER = 10 ** 6
def invoice(rng, policy):
amt = exp(rng.uniform(log(500), log(20000)))
u, i = rng.random(), 0
while u >= KINDS[i][0]:
u -= KINDS[i][0]
i += 1
_, h0, fail, kind = KINDS[i]
t2, t3 = (NEVER, NEVER) if policy == 1 else (60, 105)
credible = amt > 4762 or policy == 3
fee = 0.05 * amt if amt > 4762 else (300.0 if policy == 3
else 0.05 * amt)
cost = 0.0
for d in range(365):
if d == t2 and kind == "prompt":
cost += 0.3 * amt
if d == t3:
cost += fee + (0.3 * amt if kind == "prompt" else 0.0)
m = MULT[kind]
stage3 = d >= t3 and credible
k = m[2] if stage3 else m[1] if d >= t2 else m[0]
if rng.random() < h0 * k:
return amt, amt - cost
if rng.random() < fail:
return amt, -cost
return amt, -cost
if __name__ == "__main__":
rng = random.Random(94)
seeds = [rng.random() for _ in range(50 * 200)]
print(" policy net recovery / owed")
for p, lab in ((1, "reminders only"), (2, "timetable for all"),
(3, "timetable, credible threats")):
owed = net = 0.0
for s in seeds:
a, n = invoice(random.Random(s), p)
owed += a
net += n
print(" {:<28} {:19.4f}".format(lab, net / owed))
# policy net recovery / owed
# reminders only 0.9457
# timetable for all 0.9490
# timetable, credible threats 0.9492只发提醒,扣除成本后收回欠款的 0.9457;对每张发票使用时间表 0.9490;时间表加可信威胁 0.9492。对健康的账款组合而言,好的与平庸的催收政策只差零点几个百分点;大的损失来自错误,例如对好客户过早升级(第四章)。账款风险更高的公司,例如第三章的组合 B,从同样的纪律中获益更多。
选择开价
推导
# Chapter 8. p15: Is there a deal at all, and where to open? A seller
# whose own floor is 80 does not know the buyer's ceiling, which she
# guesses as normal with mean m and sd s. A deal is possible if the
# ceiling exceeds 80. If she opens at price x and the buyer accepts
# when x is below the ceiling (else talks fail), her expected gain over
# the floor is (x - 80) P(ceiling > x). Best opening by grid (step
# 0.5), and the cost of an overconfident guess (s 5 when the truth is
# s 15).
from math import erf, sqrt
def p_above(x, m, s):
return 0.5 * (1 - erf((x - m) / (s * sqrt(2))))
def best(m, s):
xs = [80 + i / 2 for i in range(0, 161)]
return max(xs, key=lambda x: (x - 80) * p_above(x, m, s))
if __name__ == "__main__":
print(" guess (m, s) P(deal possible) best opening expected gain")
for m, s in ((100, 5), (100, 15), (90, 15)):
x = best(m, s)
print(" ({:3d}, {:2d}) {:16.3f} {:12.1f} {:13.2f}".format(
m, s, p_above(80, m, s), x, (x - 80) * p_above(x, m, s)))
x = best(100, 5)
print(" open at {:.1f} (from s 5) when truth is s 15: gain {:.2f}"
.format(x, (x - 80) * p_above(x, 100, 15)))
# guess (m, s) P(deal possible) best opening expected gain
# (100, 5) 1.000 95.5 12.65
# (100, 15) 0.909 99.5 10.01
# ( 90, 15) 0.748 95.0 5.54
# open at 95.5 (from s 5) when truth is s 15: gain 9.58底线为 80 的卖方,猜买方的上限是 100、离散度 5,会认为一定能成交(1.000),应该开价 95.5,预期收益 12.65。离散度为 15 时,能成交的机会是 0.909,最佳开价 99.5,预期收益 10.01;如果她对上限的猜测只有 90,成交机会降到 0.748,收益降到 5.54。因为过度自信而开价 95.5、而真实的离散度较宽时,得到 9.58 而不是 10.01。这里过度自信的代价很小;看错对方区间中心的代价则很大。
从几张发票认识一个客户
推导
# Chapter 8. p16: How many invoices before a client is known? A client
# pays on time with chance r. Starting from a uniform prior, k on-time
# payments out of n give a Beta(k + 1, n - k + 1) posterior. The chance
# that r is below 0.7 (a client to put on stricter terms), and the
# posterior mean, after a few histories. Exact, by numerical
# integration on a grid of 20,000 points.
def post(k, n, cut=0.7, m=20000):
xs = [(i + 0.5) / m for i in range(m)]
w = [x ** k * (1 - x) ** (n - k) for x in xs]
z = sum(w)
below = sum(wi for x, wi in zip(xs, w) if x < cut) / z
mean = sum(x * wi for x, wi in zip(xs, w)) / z
return below, mean
if __name__ == "__main__":
print(" on time / invoices P(r < 0.7) mean r")
for k, n in ((1, 2), (3, 4), (8, 10), (16, 20), (80, 100)):
b, mu = post(k, n)
print(" {:>7} / {:<9} {:10.3f} {:6.3f}".format(k, n, b, mu))
# on time / invoices P(r < 0.7) mean r
# 1 / 2 0.784 0.500
# 3 / 4 0.528 0.667
# 8 / 10 0.313 0.750
# 16 / 20 0.198 0.773
# 80 / 100 0.014 0.794两张发票中有一张按时付款之后,客户真实按时率低于 0.7 的机会是 0.784;四张中三张之后 0.528;十张中八张之后 0.313;二十张中十六张之后 0.198;只有在一百张中八十张之后,才降到 0.014。几张发票无法给客户分类。对新客户,给予适中的条件并加以观察,好过在一次迟付之后就收紧条件;对健康的客户而言,一次迟付大多是运气不好。
利益披露
声明: 本书讨论议价、催收与升级的数学,不是法律、财务或劳动关系建议,书中参数均为示意。债务催收、征信报告与法律行动的威胁都受监管;请依照适用的法律咨询合格律师,并且绝不要以依法不能采取或并不打算采取的行动相威胁。
注
预先登记的检验: 第一章关于让步模式的假说,与第三章关于应收账款账龄的假说,都应在收集数据之前登记数据来源、定义与分析方法,结果不论正负都发表。
小结
对健康的债务人组合,主要收益来自不伤害好客户,以及让真实的威胁负担得起;开价应当建立在对对方区间的诚实估计上;不应凭几张发票评判一个客户。
Fan, Y. (2026). Multistage Ultimatum Bargaining, Accounts-Receivable Decay Kinetics and Optimal Escalation Sequences: Ackerman Concession Curves, Credible Threat Thresholds and Default Recovery Bounds [多阶段最后通牒议价、应收账款衰减动力学与最优施压序列] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090245 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).