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

第一章 多通道路由与拒付红线

先改正红线的来源与数字

常听到的说法是「发卡行的拒付红线是 0.9%」。核对之后要改正三点。第一,监控线由卡组织制定,经收单机构执行,而不是由发卡行制定。第二,0.9% 是 Visa 旧的争议监控计划(VDMP)「标准级」的门槛,而且同一个月还要至少 100 笔争议;VDMP 已于 2025 年 3 月 31 日退役。第三,2025 年 4 月 1 日起取代它的 Visa 收单机构监控计划(VAMP)计算的是欺诈报告(TC40)加争议(TC15)除以已结算的无卡交易笔数;对亚太地区商户,「过高」等级的门槛自 2026 年 4 月 1 日起为 1.5%,并且一个月要累计至少 1,500 笔才进入监控 (Visa 2025)。收单机构与支付网关往往另有更严的内部风控线。本书把 0.9% 作为商户自己的运营警戒线,把 0.65% 作为更保守的缓冲线。

把高风险订单引向即时扫码

马来西亚的买家除了刷卡,还可以用 DuitNow 二维码或电子钱包即时转账 (Bank Negara Malaysia 2018)。这类转账没有信用卡式的拒付程序,所以把风险最高的订单引导到扫码支付,就能压低卡通道的事件率。代价是一部分买家在只看到扫码选项时放弃付款。

推导

# Chapter 1. p01: Routing orders between cards and instant QR payment.
# Orders fall into five risk segments (shares 0.40/0.25/0.15/0.12/0.08,
# ticket RM 150). On the card route 90% of attempts succeed, the fee is
# 2.9% + RM 1, and a share r of settled card sales later brings a fraud
# report or dispute (0.2%, 0.4%, 0.8%, 2%, 5%); each costs RM 150 + RM 60
# with probability 0.7. A buyer offered only DuitNow/e-wallet QR (fee 1%,
# no chargebacks) completes with probability 0.85/0.82/0.78/0.70/0.60.
# Choose for each segment the share x offered cards (0, 0.25, ... 1) to
# maximise margin with the card event ratio at or below a target.
from itertools import product

SHARE = (0.40, 0.25, 0.15, 0.12, 0.08)
RATE = (0.002, 0.004, 0.008, 0.02, 0.05)
QR = (0.85, 0.82, 0.78, 0.70, 0.60)
TICKET, CARD_OK = 150.0, 0.90
GRID = (0.0, 0.25, 0.5, 0.75, 1.0)

def margins(i):
    card = CARD_OK * (TICKET - 0.029 * TICKET - 1.0 -
                      RATE[i] * 0.7 * (TICKET + 60))
    qr = QR[i] * (TICKET - 0.01 * TICKET)
    return card, qr

def evaluate(x):
    card_sales = sum(SHARE[i] * x[i] * CARD_OK for i in range(5))
    events = sum(SHARE[i] * x[i] * CARD_OK * RATE[i] for i in range(5))
    paid = card_sales + sum(SHARE[i] * (1 - x[i]) * QR[i] for i in range(5))
    m = sum(SHARE[i] * (x[i] * margins(i)[0] + (1 - x[i]) * margins(i)[1])
            for i in range(5))
    return events / card_sales, paid, m

def best(target):
    ok = [(evaluate(x)[2], x) for x in product(GRID, repeat=5)
          if sum(x) > 0 and evaluate(x)[0] <= target]
    return max(ok)[1]

if __name__ == "__main__":
    allc = (1.0,) * 5
    r, paid, m = evaluate(allc)
    print("   all on cards: event ratio {:.2%}, paid {:.3f}".format(r, paid))
    print("     RM {:,.0f} per 10,000 orders".format(1e4 * m))
    for target in (0.015, 0.009, 0.0065):
        x = best(target)
        r, paid, m = evaluate(x)
        print("   target {:.2%}: card shares {}".format(
            target, " ".join("{:.2f}".format(v) for v in x)))
        print("     ratio {:.2%}, paid {:.3f}, RM {:,.0f} per 10,000"
              " orders".format(r, paid, 1e4 * m))
#    all on cards: event ratio 0.94%, paid 0.900
#      RM 1,289,414 per 10,000 orders
#    target 1.50%: card shares 1.00 1.00 1.00 1.00 1.00
#      ratio 0.94%, paid 0.900, RM 1,289,414 per 10,000 orders
#    target 0.90%: card shares 1.00 1.00 1.00 1.00 0.75
#      ratio 0.86%, paid 0.894, RM 1,282,520 per 10,000 orders
#    target 0.65%: card shares 1.00 1.00 1.00 1.00 0.00
#      ratio 0.59%, paid 0.876, RM 1,261,838 per 10,000 orders

在示意参数下,订单按风险分成五段,事件率从 0.2% 到 5%。全部走卡时,卡通道事件率是 0.94%,每 10,000 笔订单的毛利为 RM 1,289,414。以 0.9% 为警戒线,最好的路由是让风险最高一段的四分之三走卡,事件率降到 0.86%,毛利 RM 1,282,520。以 0.65% 为缓冲线,最高一段全部改走扫码,事件率降到 0.59%,付款完成率从 0.900 降到 0.876,毛利 RM 1,261,838,约少 2%。

注

路由不能改变退款义务。 引导到扫码支付针对的是盗卡与欺诈风险;买家依法与依店规应得的退款照样要付。人为拆单或制造小额交易来稀释分母,通常违反收单协议,本书不讨论此类做法。

拒付是迟到的

监控比率的分子是当月收到的事件,分母是当月结算的交易,而一个月销售带来的事件会在之后几个月陆续到达。在示意模型中,45% 当月到达,40% 下个月到达,15% 再下个月到达。第 4、5 个月有一个欺诈团伙每月加入 1,500 笔订单,评分都落在风险最高一段。

推导

# Chapter 1. p02: Disputes arrive late. Of the fraud reports and disputes
# that a month's card sales will bring, 45% are reported that month, 40%
# the next and 15% the month after; the monitored ratio divides the
# events REPORTED in a month by the card sales SETTLED in it. 12,000
# orders a month in the segments of p01; in months 4 and 5 a fraud ring
# adds 1,500 orders a month that score in the riskiest segment, with an
# event rate of 15% if they reach cards (QR completion 5%). Routings:
# loose = the 0.90% routing of p01, strict = segments 4 and 5 half and
# none on cards. Controllers: always loose; react to LAST month's
# observed ratio (> 0.90% -> strict next month); score orders at checkout
# and go strict in any month whose expected ratio exceeds 0.90%.
# 300 simulated years each, seed 85 (counts drawn from normal or
# Poisson approximations to the binomial).
import random
from math import exp

SHARE = (0.40, 0.25, 0.15, 0.12, 0.08)
RATE = (0.002, 0.004, 0.008, 0.02, 0.05)
LOOSE = (1.0, 1.0, 1.0, 1.0, 0.75)
STRICT = (1.0, 1.0, 1.0, 0.5, 0.0)
LAG = (0.45, 0.40, 0.15)

def binom(rng, n, p):
    """Binomial draw: normal for large counts, Poisson for rare events."""
    v = n * p * (1 - p)
    if v > 30:
        return min(n, max(0, round(n * p + v ** 0.5 * rng.gauss(0, 1))))
    lam, k, u = n * p, 0, rng.random()
    term = exp(-lam)
    cum = term
    while u > cum and k < n:
        k += 1
        term *= lam / k
        cum += term
    return k

def month_orders(m):
    rows = [(12000 * s, r) for s, r in zip(SHARE, RATE)]
    attack = 1500 if m in (4, 5) else 0
    return rows, attack

def expected_ratio(m, x):
    rows, attack = month_orders(m)
    sales = sum(n * x[i] * 0.9 for i, (n, r) in enumerate(rows))
    ev = sum(n * x[i] * 0.9 * r for i, (n, r) in enumerate(rows))
    sales += attack * x[4] * 0.9
    ev += attack * x[4] * 0.9 * 0.15
    return ev / sales

def year(rng, rule):
    pipe = [0.0] * 15
    obs, prev = [], 0.0
    for m in range(1, 13):
        if rule == "loose":
            x = LOOSE
        elif rule == "react":
            x = STRICT if prev > 0.009 else LOOSE
        else:
            x = STRICT if expected_ratio(m, LOOSE) > 0.009 else LOOSE
        rows, attack = month_orders(m)
        sales = 0
        for i, (n, r) in enumerate(rows):
            k = binom(rng, int(n * x[i]), 0.9)
            sales += k
            ev = binom(rng, k, r)
            for j, w in enumerate(LAG):
                pipe[m + j] += ev * w
        k = binom(rng, int(attack * x[4]), 0.9)
        sales += k
        ev = binom(rng, k, 0.15)
        for j, w in enumerate(LAG):
            pipe[m + j] += ev * w
        prev = pipe[m] / sales
        obs.append(prev)
    return obs

if __name__ == "__main__":
    rng = random.Random(85)
    print("   controller     month: 3     4     5     6     7"
          "   months > 0.90%")
    for rule in ("loose", "react", "checkout"):
        runs = [year(rng, rule) for _ in range(300)]
        mean = [sum(r[m] for r in runs) / 300 for m in range(12)]
        over = sum(sum(1 for v in r if v > 0.009) for r in runs) / 300
        print("   {:<10} {}   {:.2f}".format(rule, " ".join(
            "{:5.2%}".format(mean[m]) for m in range(2, 7)), over))
#    controller     month: 3     4     5     6     7   months > 0.90%
#    loose      0.85% 1.37% 1.90% 1.64% 1.07%   5.32
#    react      0.86% 1.21% 1.35% 0.94% 0.67%   2.74
#    checkout   0.86% 0.76% 0.56% 0.62% 0.79%   1.33

始终按 0.9% 路由(不调整):第 4 至 7 个月的观测比率是 1.37%、1.90%、1.64%、1.07%,一年平均有 5.32 个月高于 0.9%。看上个月的比率再反应:攻击的影响要到第 5、6 个月才压下去(1.35%、0.94%),一年 2.74 个月超线。在结账时就按订单评分预测本月的事件率并立刻调整:第 4 个月就降到 0.76%,一年只有 1.33 个月超线,而这些超线主要来自正常月份的随机波动。

三种控制方式下每月观测到的卡通道事件率(p02,300 年平均)。虚线为 0.9% 与 1.5%。

小结

0.9% 是已经退役的 VDMP 标准门槛;现行的 VAMP 把欺诈报告与争议一起计算,亚太门槛是 1.5%。把风险最高的订单引向扫码,可以用约 2% 的毛利把事件率从 0.94% 压到 0.59%。拒付会迟到,看上个月比率再反应总是慢一拍;在结账时按评分预测,才能在攻击当月就控制住。

参考文献

Bank Negara Malaysia. 2018. Interoperable Credit Transfer Framework. Policy document.
Visa. 2025. Visa Acquirer Monitoring Program (VAMP): Program Overview and Merchant Thresholds. Visa Inc.

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).