5D Library › 跨境支付智能路由、拒付抗辩博弈与资金池破产边界 · English

第五章 克拉美-伦德伯格模型与滚动保证金

「零破产」买不到

把企业的现金池看作保险公司的盈余:现金从 uu 开始,每周以净毛利 cc 增加;退款批次、拒付、保证金追缴等流出冲击按泊松过程到达,每周 2 次,大小服从均值 RM 8,000 的指数分布。净毛利带一个安全附加 θ\theta:c=(1+θ)λμc = (1+\theta)\lambda\mu。这是经典的克拉美-伦德伯格模型 (Lundberg 1903; Cramér 1930),它的破产概率(现金在任何时刻跌破零)有精确解 (Asmussen and Albrecher 2010):

ψ(u)=11+θexp⁡(−θu(1+θ)μ).\psi(u) = \frac{1}{1+\theta}\exp\!\left(-\frac{\theta\,u}{(1+\theta)\mu}\right).

对任何有限的 uu,ψ(u)\psi(u) 都大于零。所以「零破产概率」是买不到的,能买的只是一个足够小的破产概率。

推导

# Chapter 5. p09: The Cramer-Lundberg model of a cash reserve. Cash
# starts at u and earns net margin c per week; outflow shocks (a batch of
# refunds, chargebacks, a reserve call) arrive as a Poisson process, 2 a
# week, with exponential sizes of mean RM 8,000. Net margin carries a
# safety loading theta = 0.25: c = (1 + theta) * 2 * 8,000. Then the
# probability of EVER running out of cash is exactly
#   psi(u) = exp(-theta u / ((1 + theta) mu)) / (1 + theta),
# which is positive for every finite u: zero ruin cannot be bought, only
# a small one. Check by simulation: 4,000 paths of 10 years, seed 85.
import random
from math import exp, log

LAM, MU, THETA = 2.0, 8000.0, 0.25
C = (1 + THETA) * LAM * MU

def psi(u, theta=THETA):
    return exp(-theta * u / ((1 + theta) * MU)) / (1 + theta)

def needed(eps, theta=THETA):
    return (1 + theta) * MU / theta * log(1 / ((1 + theta) * eps))

def simulate(rng, u, weeks=520):
    cash, t = u, 0.0
    while True:
        dt = rng.expovariate(LAM)
        t += dt
        if t > weeks:
            return False
        cash += C * dt - rng.expovariate(1 / MU)
        if cash < 0:
            return True

if __name__ == "__main__":
    rng = random.Random(85)
    print("   reserve RM   psi(u) formula   ruined within 10 years (sim)")
    for u in (0.0, 50000.0, 100000.0, 175000.0):
        sim = sum(simulate(rng, u) for _ in range(4000)) / 4000
        print("   {:10,.0f}   {:14.4f}   {:.4f}".format(u, psi(u), sim))
    for eps in (0.01, 0.001):
        print("   reserve for ruin {:.1%}: RM {:,.0f}".format(eps,
                                                          needed(eps)))
    for th in (0.10, 0.50):
        print("   with loading {:.2f}: reserve for 1% RM {:,.0f}".format(
            th, needed(0.01, th)))
#    reserve RM   psi(u) formula   ruined within 10 years (sim)
#             0           0.8000   0.8060
#        50,000           0.2292   0.2280
#       100,000           0.0657   0.0722
#       175,000           0.0101   0.0097
#    reserve for ruin 1.0%: RM 175,281
#    reserve for ruin 0.1%: RM 267,384
#    with loading 0.10: reserve for 1% RM 396,868
#    with loading 0.50: reserve for 1% RM 100,793

安全附加 0.25 时,没有储备的现金池破产概率是 0.8000(模拟十年为 0.8060);储备 RM 50,000 时是 0.2292;RM 100,000 时是 0.0657;要把破产概率压到 1.0%,需要 RM 175,281;压到 0.1%,需要 RM 267,384。储备与目标的关系是对数的:破产概率每降低十倍,只需再加一个固定的金额。安全附加对储备的影响更大:附加只有 0.10 时,1% 的目标要 RM 396,868;附加 0.50 时只要 RM 100,793。

破产概率随起始储备的变化,三种安全附加(p09,纵轴为对数刻度)。虚线为 1%。

滚动保证金:钱没丢,只是晚到

支付网关常要求「滚动保证金」:每天扣下一定比例的结算款,放满若干天后再释放。钱没有丢,但在开始扣留的那段时间里,现金流会被压低。下面模拟一家每天卡交易 RM 40,000 的商户:网关两天后付出 97%,成本占 75%,正常退款 3%;第 300 天起有 14 天退款是平常的三倍;再加上上面的随机冲击。

推导

# Chapter 5. p10: A rolling reserve meets a refund season. Card sales are
# RM 40,000 a day; the gateway pays out 97% of each day's sales (after
# fees) two days later, but holds back a share r of every day for H days
# before releasing it. Costs (goods, ads, staff) are 75% of sales and
# normal refunds 3%, paid daily; from day 300 for 14 days refunds run at
# three times normal (a post-sale return season); on top come the random
# shocks of p09 (2 a week, mean RM 8,000). The reserve is imposed from
# day 0. Cash needed so that the balance never goes below zero over two
# years in 99% of 1,000 paths = the 99th percentile of the deepest
# shortfall. Seed 85.
import random
from math import exp

def shocks(rng, lam):
    """Number of shocks in a day: Poisson(lam) by inversion."""
    k, term, u = 0, exp(-lam), rng.random()
    cum = term
    while u > cum:
        k += 1
        term *= lam / k
        cum += term
    return k

V, PAY, COST, REF = 40000.0, 0.97, 0.75, 0.03

def path(rng, r, H, days=730):
    cash, low = 0.0, 0.0
    for d in range(days):
        inflow = 0.0
        if d >= 2:
            inflow += (1 - r) * PAY * V
        if d >= H:
            inflow += r * PAY * V
        refunds = REF * V * (3 if 300 <= d < 314 else 1)
        cash += inflow - COST * V - refunds
        for _ in range(shocks(rng, 2 / 7)):
            cash -= rng.expovariate(1 / 8000.0)
        low = min(low, cash)
    return -low

def need(r, H, n=1000, seed=85):
    rng = random.Random(seed)
    s = sorted(path(rng, r, H) for _ in range(n))
    return s[int(0.99 * n) - 1]

if __name__ == "__main__":
    print("   hold r  for H days  held at peak RM  cash needed (99%) RM")
    for r, H in ((0.0, 0), (0.10, 90), (0.10, 180), (0.20, 90),
                 (0.20, 180)):
        held = r * PAY * V * H
        print("   {:6.0%}  {:10d}  {:15,.0f}  {:20,.0f}".format(
            r, H, held, need(r, H)))
#    hold r  for H days  held at peak RM  cash needed (99%) RM
#        0%           0                0               109,668
#       10%          90          349,200               154,991
#       10%         180          698,400               157,884
#       20%          90          698,400               428,954
#       20%         180        1,396,800               706,439

不扣保证金时,要让两年内 99% 的情形现金都不跌破零,需要 RM 109,668。扣 10%、放 90 天时,网关手上最多压着 RM 349,200,但商户需要的现金只多一点,RM 154,991:因为扣下 10% 之后每天的现金流仍是正的。扣 20% 时情况完全不同:每天扣下的钱多于净毛利,在开始扣留的那段时间里现金每天都在流失;放 90 天要 RM 428,954,放 180 天要 RM 706,439。

注

看保证金比例与净毛利率的大小关系。 保证金比例低于净毛利率时,它主要是一笔被压着的钱;高于净毛利率时,它在扣留期内每天都在消耗现金,所需储备随扣留天数线性增长。

小结

在克拉美-伦德伯格模型中,「零破产」不可能,储备只能把破产概率压小,而且每降低十倍只需加一个固定金额;在示意参数下,1% 需要 RM 175,281。滚动保证金是否危险,取决于它是否超过净毛利率:超过时,扣留期越长,所需储备越多。

参考文献

Asmussen, Søren, and Hansjörg Albrecher. 2010. Ruin Probabilities. 2nd ed. World Scientific.
Cramér, Harald. 1930. “On the Mathematical Theory of Risk.” In Skandia Jubilee Volume. Stockholm.
Lundberg, Filip. 1903. “I. Approximerad Framställning Af Sannolikhetsfunktionen. II. Återförsäkring Af Kollektivrisker.” PhD thesis, Uppsala University.

Fan, Y. (2026). Cross-Border Payment Routing, Chargeback Defense Games and Treasury Ruin Bounds: Multi-Acquirer Success Optimization, Dispute Evidence Scoring and Cramér-Lundberg Cash Reserves [跨境支付智能路由、拒付抗辩博弈与资金池破产边界] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090225 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).