5D Library › 多阶段最后通牒议价、应收账款衰减动力学与最优施压序列 · English

第四章 何时升级

三个阶段

常见的催收顺序是:从友好提醒,到停止服务或供货,再到律师函与向征信机构报告。每一步都会让一些债务人付得更快,也都有代价:对只是晚了一点的客户停止服务,可能失去未来的生意;律师函要花钱,也伤感情。p07 为组合 A 的债务人给这个顺序估值,并搜索第二步与第三步的最佳日期。

推导

# Chapter 4. p07: When to escalate. One overdue invoice of 1, debtor of
# book A in p05 (prompt-but-late 60%, slow 30%, distressed 10%). Daily
# hazards of paying: 1/15, 1/90, 1/400; distressed debtors also become
# insolvent at 1/250 a day. Three stages:
#   1 friendly reminders from day 0: hazards x 1.5, no cost;
#   2 stop service from day t2: x 3 for prompt and slow, x 1.5 for
#     distressed; costs 0.3 (lost future margin) if the debtor is a
#     prompt-but-late customer still unpaid on day t2;
#   3 lawyer's letter and credit reporting from day t3: x 6 for slow,
#     x 3 for distressed (who now pay this creditor first); costs 0.05
#     plus 0.3 for any prompt customer still unpaid.
# Expected net recovery within 365 days, by exact daily integration;
# grid search over t2 and t3.
KINDS = ((0.6, 1 / 15, 0.0, "prompt"), (0.3, 1 / 90, 0.0, "slow"),
         (0.1, 1 / 400, 1 / 250, "distressed"))
MULT = {"prompt": (1.5, 3.0, 3.0), "slow": (1.5, 3.0, 6.0),
        "distressed": (1.5, 1.5, 3.0)}

def value(t2, t3, days=365, parts=False):
    total, by = 0.0, []
    for share, h0, fail, kind in KINDS:
        alive, got, cost = 1.0, 0.0, 0.0
        for d in range(days):
            if d == t2 and kind == "prompt":
                cost += 0.3 * alive
            if d == t3:
                cost += 0.05 * alive + (0.3 * alive if kind == "prompt"
                                        else 0.0)
            m = MULT[kind]
            k = m[2] if d >= t3 else m[1] if d >= t2 else m[0]
            h = min(1.0, h0 * k)
            got += alive * h
            alive *= (1 - h) * (1 - fail)
        total += share * (got - cost)
        by.append(got - cost)
    return by if parts else total

if __name__ == "__main__":
    never = 10 ** 6
    grid = [(t2, t3) for t2 in range(15, 181, 15)
            for t3 in list(range(t2, 241, 15)) + [never]]
    best = max(grid, key=lambda g: value(*g))
    print("   policy                  t2     t3    net    prompt  slow"
          "   distr.")
    for lab, g in (("reminders only", (never, never)),
                   ("everything on day 1", (1, 1)),
                   ("30/60/90 habit", (60, 90)),
                   ("best on the grid", best)):
        t3 = "never" if g[1] == never else str(g[1])
        t2 = "never" if g[0] == never else str(g[0])
        p = value(*g, parts=True)
        print("   {:<21} {:>5}  {:>5}  {:6.4f}  {:5.3f}  {:5.3f}  {:5.3f}"
              .format(lab, t2, t3, value(*g), *p))
#    policy                  t2     t3    net    prompt  slow   distr.
#    reminders only        never  never  0.9450  1.000  0.998  0.456
#    everything on day 1       1      1  0.5936  0.415  0.951  0.593
#    30/60/90 habit           60     90  0.9507  0.999  0.993  0.530
#    best on the grid         60    105  0.9508  0.999  0.996  0.523

只发提醒,扣除成本后收回金额的 0.9450。第 1 天就把每一步都用上,只收回 0.5936:困境债务人多付了一点(0.593 对 0.456),但本来就会付款的及时客户流失了(0.415 而不是 1.000)。第 60 天停止服务、第 90 天发函的常见做法给出 0.9507,网格上最好的组合是第 60 与第 105 天,给出 0.9508。升级的收益是真实的,但很小,而且几乎全部来自困境债务人:他们的收回从 0.456 升到 0.523。它的代价则落在好客户身上,所以主要规则是:在他们有时间付款之前不要升级。

三种停止服务日期下,律师函日期与每张发票的净回收(p07)。

一张时间表,还是每种一张?

如果升级只对困境债务人有用,为什么不立刻对他们升级、其余的不动?p08 用一个不完美的筛选来试。

推导

# Chapter 4. p08: One timetable for all, or one for each kind? Same
# debtors and stages as p07. A uniform timetable escalates everyone at
# the same days; a segmented one escalates debtors flagged as
# distressed on day 1 and the rest on days 60 and 105. Flags come from
# a screen that marks a distressed debtor with chance s and a sound one
# with chance 0.05. Then the best uniform timetable for three costs of
# lost future margin M.
KINDS = ((0.6, 1 / 15, 0.0, "prompt"), (0.3, 1 / 90, 0.0, "slow"),
         (0.1, 1 / 400, 1 / 250, "distressed"))
MULT = {"prompt": (1.5, 3.0, 3.0), "slow": (1.5, 3.0, 6.0),
        "distressed": (1.5, 1.5, 3.0)}

def one(kind, h0, fail, t2, t3, margin=0.3, days=365):
    alive, got, cost = 1.0, 0.0, 0.0
    for d in range(days):
        if d == t2 and kind == "prompt":
            cost += margin * alive
        if d == t3:
            cost += 0.05 * alive + (margin * alive if kind == "prompt"
                                    else 0.0)
        m = MULT[kind]
        k = m[2] if d >= t3 else m[1] if d >= t2 else m[0]
        h = min(1.0, h0 * k)
        got += alive * h
        alive *= (1 - h) * (1 - fail)
    return got - cost

def segmented(s, false_flag=0.05):
    total = 0.0
    for share, h0, fail, kind in KINDS:
        flag = s if kind == "distressed" else false_flag
        v = flag * one(kind, h0, fail, 1, 1) + \
            (1 - flag) * one(kind, h0, fail, 60, 105)
        total += share * v
    return total

def uniform(t2, t3, margin):
    return sum(sh * one(k, h, f, t2, t3, margin) for sh, h, f, k in KINDS)

if __name__ == "__main__":
    print("   uniform 60 / 105: {:.4f}".format(uniform(60, 105, 0.3)))
    print("   screen catches s of distressed   segmented value")
    for s in (0.0, 0.5, 0.8, 1.0):
        print("   {:31.1f}   {:15.4f}".format(s, segmented(s)))
    print("   perfect screen (s = 1, no false flags): {:.4f}".format(
        segmented(1.0, 0.0)))
    print("   lost margin M   best uniform t2   t3    value")
    grid = [(a, b) for a in range(15, 181, 15) for b in range(a, 241, 15)]
    for m in (0.1, 0.3, 0.6):
        g = max(grid, key=lambda x: uniform(x[0], x[1], m))
        print("   {:13.1f}   {:15d}   {:3d}   {:6.4f}".format(
            m, g[0], g[1], uniform(g[0], g[1], m)))
#    uniform 60 / 105: 0.9508
#    screen catches s of distressed   segmented value
#                                0.0            0.9326
#                                0.5            0.9361
#                                0.8            0.9382
#                                1.0            0.9396
#    perfect screen (s = 1, no false flags): 0.9578
#    lost margin M   best uniform t2   t3    value
#              0.1                60   105   0.9510
#              0.3                60   105   0.9508
#              0.6                75   105   0.9506

统一的时间表给出 0.9508。对被标记为困境的债务人立刻升级,如果筛选一个也没抓到,给出 0.9326;即使全部抓到,也只有 0.9396,因为它同时把 5% 的健康客户也标记了,而把好客户当坏客户对待的代价,大于从困境债务人多收回的部分。只有完美的、没有误标的筛选才更好:0.9578。最佳的统一日期几乎不随失去的未来利润而变化(0.9510、0.9508 与 0.9506)。教训与第 093 卷关于早期预警的一章相同:当所要寻找的情况罕见时,即使误报率很小的筛选,代价也可能超过它节省的。

假说 · 尚未验证

作者的诠释(尚未验证): 对逾期账款以「只是晚了」的客户为主的债权人而言,把第一个强制性步骤(停止服务)提前到大约两个月之前,会降低计入流失客户价值后的净回收。什么会证明它错: 在一项随机分配第一个强制性步骤日期的预先登记现场实验中,较早的步骤提高了计入客户终身价值后的净回收。

小结

升级只对本来不会付款的债务人有效,而它的代价落在本来会付款的客户身上。对所有人晚一点升级,胜过早升级;只有在筛选很少误标好客户时,针对性升级才值得。

Fan, Y. (2026). Multistage Ultimatum Bargaining, Accounts-Receivable Decay Kinetics and Optimal Escalation Sequences: Ackerman Concession Curves, Credible Threat Thresholds and Default Recovery Bounds [多阶段最后通牒议价、应收账款衰减动力学与最优施压序列] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090245 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).