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

第二章 比率是带噪声的计数,退款要快于拒付

小商户的比率会自己跳

一个月的事件数是一个随机计数。真实事件率 0.60% 的商户,每月结算的卡交易越少,观测比率越容易偶然越线。把每月事件数看作泊松分布,可以算出偶然越线的概率。

推导

# Chapter 2. p03: A ratio is a noisy count. A merchant whose true event
# rate is 0.60% (or 0.75%) settles N card sales a month; the monthly count
# of events is Poisson with mean rate * N. How often does the observed
# ratio exceed 0.90% (the old Visa VDMP "standard" line, which also needed
# at least 100 disputes) or 1.50% (the VAMP "excessive" line for Asia
# Pacific merchants from April 2026, which needs 1,500 events a month)?
from math import ceil, exp, lgamma, log

def tail(mean, k):
    """P(X >= k) for X ~ Poisson(mean), summed upward in log space."""
    total, i = 0.0, k
    while True:
        t = exp(i * log(mean) - mean - lgamma(i + 1))
        total += t
        if i > mean and t < 1e-15:
            return min(1.0, total)
        i += 1

if __name__ == "__main__":
    print("   card sales   P(ratio > 0.90%)        P(ratio > 1.50%)")
    print("   per month    true 0.60%  true 0.75%  true 0.75%")
    for n in (2000, 5000, 20000, 100000):
        k9 = int(0.009 * n) + 1
        k15 = int(0.015 * n) + 1
        print("   {:10,d}   {:10.4f}  {:10.4f}  {:10.4f}".format(
            n, tail(0.006 * n, k9), tail(0.0075 * n, k9),
            tail(0.0075 * n, k15)))
    for need, line in ((100, 0.009), (1500, 0.015)):
        print("   floor of {:,} events: hit at {:.2%} from {:,} sales/month"
              .format(need, line, ceil(need / line)))
#    card sales   P(ratio > 0.90%)        P(ratio > 1.50%)
#    per month    true 0.60%  true 0.75%  true 0.75%
#         2,000       0.0374      0.1805      0.0002
#         5,000       0.0040      0.0985      0.0000
#        20,000       0.0000      0.0076      0.0000
#       100,000       0.0000      0.0000      0.0000
#    floor of 100 events: hit at 0.90% from 11,112 sales/month
#    floor of 1,500 events: hit at 1.50% from 100,000 sales/month

每月 2,000 笔卡交易时,真实事件率 0.60% 的商户有 0.0374 的概率在某个月观测到高于 0.9% 的比率;真实事件率 0.75% 时,这个概率是 0.1805。每月 20,000 笔时,后者降到 0.0076。旧的 VDMP 要求同时有至少 100 笔争议,在 0.9% 的比率下这要每月 11,112 笔卡交易才达到;VAMP 在亚太要求至少 1,500 笔事件,在 1.5% 的比率下要每月 100,000 笔卡交易才达到。

注

小商户不在卡组织的正式监控范围内,并不等于安全。 计数门槛使小商户很少被卡组织正式列入监控,但收单机构与网关通常按自己的比率与笔数规则冻结或收取保证金。第五、六章计算这类冻结对现金的影响。

退款要快于拒付

许多拒付来自本来可以退款的买家:他们申请退款,等不到回应,就去银行争议。下面假设旺季一个月有 600 位买家申请退款;等待 tt 天仍未退款的买家,会以 0.5(1−e−t/5)0.5(1-e^{-t/5}) 的概率发起争议。争议几乎都判买家胜(「未处理退款」),商户照样损失 RM 150,另付 RM 60 的手续费,而且这笔争议计入监控比率。

推导

# Chapter 2. p04: Refund before they dispute. In a peak month 600 of
# 12,000 buyers (10,800 settled on cards) ask for their money back. A
# buyer still waiting t days after asking files a dispute with
# probability 0.5 * (1 - exp(-t / 5)). A dispute almost always goes the
# buyer's way ("credit not processed"): the merchant loses the RM 150
# anyway, plus a RM 60 fee, and the dispute counts in the monitored ratio.
from math import exp

ASK, CARD, TICKET, FEE = 600, 10800, 150.0, 60.0

def month(t):
    d = ASK * 0.5 * (1 - exp(-t / 5.0))
    cost = ASK * TICKET + d * FEE
    return d, d / CARD, cost

if __name__ == "__main__":
    print("   refund within (days)  disputes  added to ratio  cost RM")
    for t in (1, 3, 7, 14):
        d, r, c = month(t)
        print("   {:20d}  {:8.0f}  {:14.2%}  {:7,.0f}".format(t, d, r, c))
    d1, d14 = month(1)[0], month(14)[0]
    print("   refunding in 1 day instead of 14 avoids {:.0f} disputes"
          .format(d14 - d1))
#    refund within (days)  disputes  added to ratio  cost RM
#                       1        54           0.50%   93,263
#                       3       135           1.25%   98,121
#                       7       226           2.09%  103,561
#                      14       282           2.61%  106,905
#    refunding in 1 day instead of 14 avoids 227 disputes

1 天内退款时,旺季这个月仍有 54 笔争议,使比率增加 0.50%;7 天内退款时有 226 笔,增加 2.09%;14 天时 282 笔,增加 2.61%。从 14 天缩短到 1 天,可以避免 227 笔争议。就现金而言差别不大(RM 93,263 对 RM 106,905),但对比率而言,慢退款本身就足以把商户推过任何一条红线。

假说 · 尚未验证

作者的假说(尚未验证): 对数字产品商户,把退款承诺从「7 个工作日」改为「24 小时内处理」,会使下一个季度的争议笔数至少减少三成。什么会证明它错: 在按订单随机分配退款时限的试验中,争议率的差距小于三成。

小结

比率是带噪声的计数:每月卡交易越少,越容易偶然越线,而卡组织的笔数门槛使小商户很少进入正式监控,但收单机构有自己的规则。最便宜的拒付控制是快速退款:在示意的旺季中,1 天内退款比 14 天少 227 笔争议。

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).