5D Library › 股东搭便车博弈、动态股权成熟与背刺早期预警模型 · English

第五章 用沃尔德序贯检验做早期预警

信号与检验

准备离开并另起炉灶的合伙人很少会宣布。可能显露的是一种模式:关键资源一点一点、每次都说得过去地移出公司,例如一个客户关系转到私人渠道,代码放进私人仓库,一份供应商合同改以别人的名义重签。每一周要么发生这样的事,要么不发生。如果这类事件平时很少(每周 0.02),在背叛前的几个月里较多(每周 0.10),要不要拉响警报,就是在两个假设之间做检验。

沃尔德的序贯概率比检验平均用最少的观察来回答这个问题 (Wald 1945)。每过一周,它把似然比的对数加进累计和,当累计和越过由容许误差率决定的两条边界之一时就停止。

推导

# Chapter 5. p09: Wald's sequential probability ratio test. Each week a
# partner either does or does not move a key resource outside the firm
# without an agreed reason (a client moved to a private channel, code
# kept in a private repository, a supplier contract re-signed in
# another name). Normal rate p0 = 0.02 a week; rate when preparing to
# defect p1 = 0.10. Error targets alpha = 0.01 (false alarm) and
# beta = 0.05 (miss). Wald's bounds: A = (1 - beta) / alpha and
# B = beta / (1 - alpha) on the likelihood ratio. Simulated error rates
# and average weeks to a decision, against the shortest fixed-length
# test with the same error targets. 20,000 runs each, seed 93.
import random
from math import log

P0, P1, ALPHA, BETA = 0.02, 0.10, 0.01, 0.05
UP, DOWN = log((1 - BETA) / ALPHA), log(BETA / (1 - ALPHA))
YES, NO = log(P1 / P0), log((1 - P1) / (1 - P0))

def binom_tail(n, k, p):
    # P(X >= k) for X ~ binomial(n, p)
    c, tot = 1.0, 0.0
    for i in range(n + 1):
        if i >= k:
            tot += c * p ** i * (1 - p) ** (n - i)
        c = c * (n - i) / (i + 1)
    return tot

def fixed_test():
    for n in range(1, 400):
        for k in range(n + 1):
            if binom_tail(n, k, P0) <= ALPHA:
                if 1 - binom_tail(n, k, P1) <= BETA:
                    return n, k
                break
    return None

def sprt(rng, p, cap=520):
    s = 0.0
    for week in range(1, cap + 1):
        s += YES if rng.random() < p else NO
        if s >= UP:
            return "alarm", week
        if s <= DOWN:
            return "clear", week
    return "none", cap

if __name__ == "__main__":
    rng = random.Random(93)
    print("   bounds on log LR: {:.3f} and {:.3f}".format(UP, DOWN))
    print("   step per event {:+.3f}; per quiet week {:+.4f}".format(YES, NO))
    print("   events needed with no quiet weeks: {:.1f}".format(UP / YES))
    for lab, p in (("normal partner", P0), ("defecting partner", P1)):
        res = [sprt(rng, p) for _ in range(20000)]
        al = sum(1 for r in res if r[0] == "alarm") / len(res)
        wk = sum(r[1] for r in res) / len(res)
        print("   {:<18} alarm rate {:.4f}   mean weeks {:5.1f}".format(
            lab, al, wk))
    n, k = fixed_test()
    print("   fixed test with the same targets: {} weeks, alarm at {}"
          " events".format(n, k))
#    bounds on log LR: 4.554 and -2.986
#    step per event +1.609; per quiet week -0.0852
#    events needed with no quiet weeks: 2.8
#    normal partner     alarm rate 0.0055   mean weeks  57.8
#    defecting partner  alarm rate 0.9534   mean weeks  56.8
#    fixed test with the same targets: 116 weeks, alarm at 7 events

误报目标 0.01、漏报目标 0.05 时,两条边界是 4.554 与 -2.986。每次事件加 1.609,每个平静的周减 0.0852;如果中间没有平静的周,2.8 次事件就够了。模拟中,这个检验对正常合伙人误报 0.0055,对背叛的合伙人发现 0.9534,平均分别在 57.8 与 56.8 周后作出决定。达到同样误差目标的固定长度检验需要 116 周。序贯检验把等待时间减半,但一年的观察仍然是一年:这类预警看得见缓慢的准备,看不见突然的行动。

两位正常合伙人(绿)与两位背叛合伙人(红)的似然比对数累计,以及两条边界(p09)。

警报中有多少是真的?

推导

# Chapter 5. p10: How many alarms are true? Suppose 2% of partners in a
# year are actually preparing to defect. A test with false-alarm rate
# alpha and detection rate 1 - beta raises alarms on
#   0.02 (1 - beta) + 0.98 alpha
# of partners, of whom the true share (positive predictive value) is
#   0.02 (1 - beta) / (0.02 (1 - beta) + 0.98 alpha).
# Then a firm watches 10 partners each year: the chance of at least one
# false alarm over five years (50 partner-years).
if __name__ == "__main__":
    print("   alpha    1 - beta   true share of alarms   P(false alarm,"
          " 10 x 5 yrs)")
    for alpha, power in ((0.01, 0.95), (0.05, 0.95), (0.001, 0.80),
                         (0.10, 0.99)):
        tp = 0.02 * power
        fp = 0.98 * alpha
        ppv = tp / (tp + fp)
        five = 1 - (1 - alpha) ** 50
        print("   {:5.3f}   {:8.2f}   {:20.3f}   {:28.3f}".format(
            alpha, power, ppv, five))
    print("   base rate 10% instead, alpha 0.01: {:.3f}".format(
        0.10 * 0.95 / (0.10 * 0.95 + 0.90 * 0.01)))
#    alpha    1 - beta   true share of alarms   P(false alarm, 10 x 5 yrs)
#    0.010       0.95                  0.660                          0.395
#    0.050       0.95                  0.279                          0.923
#    0.001       0.80                  0.942                          0.049
#    0.100       0.99                  0.168                          0.995
#    base rate 10% instead, alpha 0.01: 0.913

背叛是罕见的。如果一年里有 2% 的合伙人在准备背叛,误报率 0.01、发现率 0.95 的检验,其警报中 0.660 是真的;误报率 0.05 时只有 0.279。用第一种检验观察十位合伙人五年,至少有一次误报的机会是 0.395,用第二种是 0.923。非常严格的检验(误报 0.001、发现 0.80)把真警报的比例提高到 0.942,五年内误报的机会降到 0.049。如果基础比率是 10%,第一种检验的警报有 0.913 是真的。这道算术并不禁止早期预警;它禁止的是把警报当作判决。

假说 · 尚未验证

作者的诠释(尚未验证): 高管离职另起炉灶之前的几个月里,关键资源无合理说明的每周转移率会有可测量的上升。什么会证明它错: 在一项使用依法收集并匿名化的事件日志的预先登记研究中,竞争性离职前几个月的这类转移率,与留任高管没有差别。

注

法律与隐私。 对员工或合伙人的任何监测都必须合法、相称,在法律要求时事先告知,并且限于公司的业务系统。警报应当开启一个让当事人可以解释的公平调查,而绝不是一项指控。

小结

沃尔德检验发现资源转移率的缓慢变化,比同样误差目标的固定检验快约一倍。由于背叛罕见,除非检验非常严格,警报中会有很大一部分是假的,所以警报应当开启调查,而不是结束调查。

参考文献

Wald, Abraham. 1945. “Sequential Tests of Statistical Hypotheses.” The Annals of Mathematical Statistics 16 (2): 117 to 186.

Fan, Y. (2026). Shareholder Free-Riding Games, Dynamic Vesting Mechanisms and Defection Early-Warning Models: Moral-Hazard Equity Dilution, Shotgun Buy-Sell Clauses and Behavioral Divergence Signals [股东搭便车博弈、动态股权成熟与背刺早期预警模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090243 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).