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
第三章 为什么逾期发票看起来每个月都更难收回
账龄表
催收机构与信用经理之间流传着一些表格,显示逾期债务收回的机会随账龄急剧下降,常附有「90 天后跌破 40%」这类数字。本书未能把这类表格追溯到一项注明数据的原始研究,因此把这个数字当作待检验的说法。至少,这种形状很容易造出来。p05 用三种债务人构造两组各 10,000 张逾期发票:只是晚了一点的及时付款人、慢付款人,以及可能破产的困境债务人。关键在于:没有任何人自己付款的机会随时间改变。
推导
# Chapter 3. p05: Why overdue invoices look less collectible every
# month. 10,000 overdue invoices of equal size. Debtors are of three
# kinds: prompt-but-late (pay with hazard 1/15 a day), slow (1/90 a
# day) and distressed (1/400 a day, and each day a 1/250 chance of
# becoming insolvent, after which nothing is paid). Nobody's own hazard
# changes with time. Book A: 60% / 30% / 10%. Book B, a riskier book:
# 40% / 30% / 30%. For invoices still unpaid on day t, the share
# eventually collected within two years. Seed 94.
import random
RATES = ((1 / 15, 0.0), (1 / 90, 0.0), (1 / 400, 1 / 250))
def invoice(rng, mix):
u, k = rng.random(), 0
while u >= mix[k]:
u -= mix[k]
k += 1
pay, fail = RATES[k]
t_pay = rng.expovariate(pay)
t_fail = rng.expovariate(fail) if fail else 1e9
return t_pay if t_pay < min(t_fail, 730) else None
def table(rng, mix):
inv = [invoice(rng, mix) for _ in range(10000)]
rows = []
for t in (0, 30, 60, 90, 180, 365):
open_ = [x for x in inv if x is None or x > t]
col = sum(1 for x in open_ if x is not None) / len(open_)
rows.append((t, len(open_) / len(inv), col))
return rows
if __name__ == "__main__":
rng = random.Random(94)
a = table(rng, (0.6, 0.3, 0.1))
b = table(rng, (0.4, 0.3, 0.3))
print(" days book A: unpaid collectible book B: unpaid"
" collectible")
for (t, ua, ca), (_, ub, cb) in zip(a, b):
print(" {:4d} {:14.3f} {:11.3f} {:14.3f} {:11.3f}".format(
t, ua, ca, ub, cb))
# days book A: unpaid collectible book B: unpaid collectible
# 0 1.000 0.939 1.000 0.816
# 30 0.391 0.845 0.543 0.660
# 60 0.258 0.765 0.418 0.559
# 90 0.200 0.697 0.353 0.478
# 180 0.114 0.469 0.256 0.279
# 365 0.068 0.111 0.199 0.071在组合 A 里,全部逾期发票最终有 0.939 收回。但在第 90 天仍未付的发票中,只有 0.697 会收回;第 180 天 0.469;第 365 天 0.111。在风险更高的组合 B 里,开始时是 0.816,第 90 天 0.478,第 180 天 0.279。账龄曲线急剧下降,而「90 天跌破 40%」需要风险更高的组合。然而没有任何债务人的付款倾向下降过:到第 90 天,及时付款人已经付清离开,仍未付的发票大多属于慢付款人与困境债务人。组合 A 在第 90 天仍未付的只有 0.200,组合 B 是 0.353。
筛选还是衰减?
这是否意味着时间无关紧要?p06 只看慢付款人,把两种效应分开。
推导
# Chapter 3. p06: Selection or decay? Take the slow debtors of p05
# alone (hazard of paying 1/90 a day). Case A: their hazard stays
# constant. Case B: each debtor's hazard itself decays, halving every
# 60 days (memories fade, records are lost, the relationship cools).
# Case C: as A, but with a 1/250 daily chance of insolvency. For an
# invoice still unpaid on day t, the chance it is ever paid within two
# years, computed from the hazards (daily steps).
from math import exp, log
def collect_from(t, hazard, fail=0.0, end=730):
alive, got = 1.0, 0.0
for d in range(t, end):
h = hazard(d)
got += alive * h
alive *= (1 - h) * (1 - fail)
return got
if __name__ == "__main__":
a = lambda d: 1 / 90
b = lambda d: (1 / 90) * exp(-log(2) * d / 60)
print(" day t A constant B own decay C insolvency")
for t in (0, 30, 90, 180):
print(" {:5d} {:10.3f} {:11.3f} {:12.3f}".format(
t, collect_from(t, a), collect_from(t, b),
collect_from(t, a, 1 / 250)))
# day t A constant B own decay C insolvency
# 0 1.000 0.621 0.737
# 30 1.000 0.496 0.737
# 90 0.999 0.290 0.737
# 180 0.998 0.114 0.737如果慢付款人的风险率不变,两年内收回的机会在第 0 天是 1.000,第 180 天仍有 0.998:等待几乎没有代价。如果债务人自己的倾向每 60 天减半,它从 0.621 降到第 30 天的 0.496、第 90 天的 0.290、第 180 天的 0.114:这时时间真的在摧毁价值。不变的破产风险,无论哪一天都给出 0.737,因为不随时间变化的风险不会让旧发票比新发票更差。账龄表分不出这几种情形;只有债务人自己的行为分得出来。对催收者而言,有用的问题不是「这张发票多老了?」,而是「这是哪一种债务人,他们的付款能力或意愿是否在下降?」
假说 · 尚未验证
作者的诠释(尚未验证): 收回率随债务账龄的下降,大部分反映的是仍未付款的债务人组成的变化,而不是每个债务人自身付款倾向的下降。什么会证明它错: 在一项对附有债务人特征的债权人记录的预先登记分析中,把债务人类型固定之后,债务人内部的付款风险率随账龄下降的幅度,与合并的收回率相当。
小结
即使没有任何债务人改变,账龄曲线也会下降,因为早付款的债务人最先离开。流行的「90 天 40%」取决于债务人组合;对行动有意义的是某个债务人的能力或意愿是否在下降。
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).