关于「滞后:同期群口径、到达日口径与暴涨时的低估」,《收单风控阶跃与准备金生存》给出的定量结论是什么?

结论:第二章证明拒付晚到使按到达日计算的比率在暴涨期最低只剩真实比率的 35.5 %,一次二十天的活动之后又冲到 1.2622 %,并用报告延迟调整把估计拉回真实值。

数据口径:书中参数为示意,全部计算由书中程序生成。 本书是非个人化的经营与风险分析,不构成法律、融资或投资建议,不向任何人承诺收入。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 35.5

mean lag 36.2 days; P(lag <= 30) 0.4575; P(lag <= 60) 0.8778
daypaymentsarrival 30dprior-monthcohort
601,0000.5000%0.5000%0.5000%
753,0000.3290%0.5045%0.5000%
905,0000.1903%0.5835%0.5000%
1055,0000.1940%0.5734%0.5000%
1205,0000.2767%0.4512%0.5000%
1355,0000.3753%0.4139%0.5000%
1505,0000.4405%0.4405%0.5000%
1655,0000.4747%0.4747%0.5000%
1805,0000.4904%0.4904%0.5000%
lowest arrival 30d ratio 0.1774% on day 96; it reads 35.5% of the true rate

核心公式

\[\hat p = \frac{\text{已到达拒付}}{\sum_c n_c\,F(a_c)}.\]
查看完整验算程序
# Chapter 2. p21: arrival-day ratios understate the true rate in a surge.
# Every payment has the same true dispute rate P = 0.5 %, but a dispute
# arrives d days after the payment, d following a discretised Erlang law
# (shape 3, scale 12 days, mean about 36 days, cut at 120 days).
# Volume: 1,000 payments a day for 60 days, a linear ramp to 5,000 a
# day over days 61 to 90, then 5,000 a day to day 180.
# Three ratios, all from expected counts (no noise):
#   arrival 30d = disputes arriving in the last 30 days / payments in
#                 the last 30 days (the ratio most dashboards show);
#   prior-month = disputes arriving in a 30-day month / payments in the
#                 month before (a lagged-denominator ratio);
#   cohort      = disputes eventually charged to a payment day / its
#                 payments, which is P by construction.
# All numbers illustrative.
import math

P = 0.005
SHAPE, SCALE, CUT = 3, 12.0, 120
DAYS = 180


def erlang_cdf(x):
    if x <= 0:
        return 0.0
    y = x / SCALE
    terms = [y ** i / math.factorial(i) for i in range(SHAPE)]
    return 1.0 - math.exp(-y) * math.fsum(terms)


def lag_pmf():
    raw = [erlang_cdf(d) - erlang_cdf(d - 1) for d in range(1, CUT + 1)]
    tot = math.fsum(raw)
    return [r / tot for r in raw]


def volume(t):
    if t <= 60:
        return 1000.0
    if t <= 90:
        return 1000.0 + 4000.0 * (t - 60) / 30.0
    return 5000.0


def arrivals(pmf):
    """expected disputes arriving on days 1..DAYS; the shop ran at
    1,000 a day for a long time before day 1"""
    out = []
    for t in range(1, DAYS + 1):
        terms = [volume(t - d) * P * pmf[d - 1] for d in range(1, CUT + 1)]
        out.append(math.fsum(terms))
    return out


def main():
    pmf = lag_pmf()
    mean_lag = math.fsum([d * q for d, q in zip(range(1, CUT + 1), pmf)])
    vol = [volume(t) for t in range(1, DAYS + 1)]
    arr = arrivals(pmf)
    print("mean lag {:.1f} days; P(lag <= 30) {:.4f}; P(lag <= 60) "
          "{:.4f}".format(mean_lag, math.fsum(pmf[:30]),
                          math.fsum(pmf[:60])))
    print(" day   payments  arrival 30d  prior-month   cohort")
    worst = (1.0, 0)
    for t in range(60, DAYS + 1):
        a30 = math.fsum(arr[t - 30:t]) / math.fsum(vol[t - 30:t])
        if a30 < worst[0]:
            worst = (a30, t)
        if t % 15 == 0:
            pm = math.fsum(arr[t - 30:t]) / math.fsum(vol[t - 60:t - 30])
            print("{:>4}  {:>9,.0f}    {:.4%}     {:.4%}    {:.4%}".format(
                t, vol[t - 1], a30, pm, P))
    print("lowest arrival 30d ratio {:.4%} on day {}; it reads {:.1%}"
          " of the true rate".format(worst[0], worst[1], worst[0] / P))


if __name__ == "__main__":
    main()

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《数字商品收单风控阶跃模型、多网关热备切换与防冻卡滚动准备金母典》第 2 章「滞后:同期群口径、到达日口径与暴涨时的低估」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-397 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026397,
  author    = {Fan, Yafu},
  title     = {Payment-Acquirer Risk Step-Functions, Multi-Gateway Hot Failover and Rolling-Reserve Survival for Digital Fleets},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论