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

第三章 友好欺诈的抗辩博弈与证据评分

白嫖者算的是什么账

「友好欺诈」指买家收到了商品,却向银行声称「未授权」或「未收到」,以拿回货款。它是一个博弈:买家决定是否争议,商户决定是否抗辩(提交证据,称为再陈述),银行按证据裁决。买家在争议之前会(多半是直觉地)估计商户抗辩胜诉的概率 ww。这是一类检查博弈 (Avenhaus et al. 2002)。

在示意模型中,商品是一份 RM 150 的数字报告。商户抗辩胜诉时,买家要退回款项,并失去账户(对他而言价值 RM 50)。买家发起争议的麻烦成本在 RM 0 到 RM 120 之间均匀分布;当 150(1−w)−50w150(1-w) – 50w 大于这个麻烦成本时,他就争议。商户在 150w150w 大于准备一份抗辩材料的成本时才抗辩:人工整理 RM 60,自动生成证据包 RM 15。每 10,000 笔订单中有 3% 的买家会动这个念头。

推导

# Chapter 3. p05: Friendly fraud as a game. A buyer who received the
# product (a RM 150 digital report) can falsely claim "not authorised" or
# "not received". If the merchant fights the dispute with evidence and
# wins (probability w), the buyer pays the money back and loses access to
# the account, a cost to him of RM 50. Buyers differ in the hassle cost h
# of disputing, uniform on RM 0 to 120. A buyer disputes when
#   150 * (1 - w) - 50 * w > h.
# The merchant fights when w * 150 exceeds the cost of preparing the
# case: RM 60 by hand, RM 15 with an automatic evidence package.
# 3% of buyers are tempted; per 10,000 orders.
A, BAN, HMAX, TEMPTED = 150.0, 50.0, 120.0, 0.03

def dispute_share(w):
    gain = A * (1 - w) - BAN * w
    return min(1.0, max(0.0, gain / HMAX))

def loss(w, prep):
    q = dispute_share(w)
    n = 10000 * TEMPTED * q
    if w * A > prep:
        return q, n, n * (prep + (1 - w) * A)
    return q, n, n * A

if __name__ == "__main__":
    print("   win prob  share who dispute  false disputes  loss RM"
          " (by hand / auto)")
    for w in (0.2, 0.35, 0.5, 0.6, 0.75):
        q, n, l1 = loss(w, 60.0)
        _, _, l2 = loss(w, 15.0)
        print("   {:8.2f}  {:17.3f}  {:14.0f}  {:7,.0f} / {:,.0f}".format(
            w, q, n, l1, l2))
    w0 = A / (A + BAN)
    print("   no tempted buyer disputes once w >= {:.2f}".format(w0))
#    win prob  share who dispute  false disputes  loss RM (by hand / auto)
#        0.20              0.917             275   41,250 / 37,125
#        0.35              0.667             200   30,000 / 22,500
#        0.50              0.417             125   16,875 / 11,250
#        0.60              0.250              75    9,000 / 5,625
#        0.75              0.000               0        0 / 0
#    no tempted buyer disputes once w >= 0.75

胜诉概率只有 0.20 时,动念头的买家中有 0.917 会争议,每 10,000 笔订单 275 笔虚假争议,人工整理时商户干脆不抗辩,损失 RM 41,250;自动证据包使抗辩划算,损失降到 RM 37,125。胜诉概率 0.50 时,争议者降到 0.417,损失 RM 16,875 或 RM 11,250;0.60 时降到 0.250。胜诉概率达到 0.75 时,再也没有动念头的买家会争议:证据的威慑作用在争议发生之前就生效了。

动念头的买家中发起争议的比例随商户胜诉概率的变化(p05)。

证据包怎样提高胜诉概率

推导

# Chapter 3. p06: Scoring an evidence package. Illustrative logistic
# model of winning a "fraud" dispute (Visa reason code 10.4): log-odds of
# a win = -1.2 plus the weight of each element present:
#   delivery/download log with timestamps        +0.6
#   terms-accepted tick, with version and time   +0.3
#   IP or device matches the account's history   +0.7
#   messages from the buyer after delivery       +0.5
# Visa Compelling Evidence 3.0 (from April 2023): two earlier undisputed
# card purchases 120 to 365 days old whose data match (two of user ID,
# IP, delivery address, device ID; one of them IP or device) shift the
# liability for a 10.4 dispute; modelled here as a win probability 0.95.
from math import exp

W = {"log": 0.6, "terms": 0.3, "ip/device": 0.7, "messages": 0.5}

def win(items, ce3=False):
    if ce3:
        return 0.95
    z = -1.2 + sum(W[i] for i in items)
    return 1 / (1 + exp(-z))

if __name__ == "__main__":
    pk = ((), ("terms",), ("log",), ("log", "terms"),
          ("log", "terms", "ip/device"),
          ("log", "terms", "ip/device", "messages"))
    for items in pk:
        print("   {:<34} win {:.3f}".format(" + ".join(items) or "nothing",
                                            win(items)))
    print("   {:<34} win {:.3f}".format("CE 3.0 match (10.4 only)",
                                        win((), True)))
    full = ("log", "terms", "ip/device", "messages")
    for i in full:
        rest = tuple(x for x in full if x != i)
        print("   dropping {:<10} costs {:.3f} of win probability".format(
            i, win(full) - win(rest)))
#    nothing                            win 0.231
#    terms                              win 0.289
#    log                                win 0.354
#    log + terms                        win 0.426
#    log + terms + ip/device            win 0.599
#    log + terms + ip/device + messages win 0.711
#    CE 3.0 match (10.4 only)           win 0.950
#    dropping log        costs 0.137 of win probability
#    dropping terms      costs 0.065 of win probability
#    dropping ip/device  costs 0.161 of win probability
#    dropping messages   costs 0.112 of win probability

下面用一个示意的逻辑模型给证据包评分,对象是「欺诈类」争议(Visa 原因代码 10.4)。什么都不交,胜诉概率 0.231;只交勾选条款的记录,0.289;交带时间戳的交付或下载日志,0.354;日志加条款,0.426;再加上 IP 或设备与账户历史相符,0.599;再加上买家在交付后发来的消息,0.711。四项中,去掉 IP 或设备一项损失最大(0.161),去掉条款勾选记录损失最小(0.065)。

Visa 的「强有力证据 3.0」(2023 年 4 月起)对原因代码 10.4 的争议另有规定:同一张卡在争议前 120 至 365 天内有两笔未被争议的交易,且其数据与争议交易相符(用户编号、IP、收货地址、设备编号中至少两项,其中一项必须是 IP 或设备),责任即转移给发卡行 (Visa 2023)。模型中把这种情形记为胜诉概率 0.950。

注

「理性消费勾选」的分量并不大。 在示意模型中,条款勾选记录是四项证据里最弱的一项;银行更看重的是交付与使用的客观记录,以及与账户历史相符的身份数据。勾选记录仍然值得保留,但不能代替日志。

小结

友好欺诈是一个威慑博弈:商户的胜诉概率越高,动念头的买家越少争议,胜诉概率达到 0.75 时虚假争议在示意模型中完全消失。证据包中最有分量的是交付日志与相符的 IP 或设备;符合强有力证据 3.0 的匹配交易可以转移欺诈类争议的责任。

参考文献

Avenhaus, Rudolf, Bernhard von Stengel, and Shmuel Zamir. 2002. “Inspection Games.” Chap. 51 in Handbook of Game Theory with Economic Applications, edited by Robert J. Aumann and Sergiu Hart, vol. 3. Elsevier.
Visa. 2023. Visa Core Rules: Compelling Evidence 3.0 for Dispute Condition 10.4 (Other Fraud, Card-Absent Environment). 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).