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

第四章 沙普利值与贡献的漂移

每位创始人加了什么

如果股权应当跟着贡献走,就必须定义贡献。沙普利值为任何一种「群体可以创造价值」的博弈下了定义:它把每个参与者在所有可能的到场顺序中的边际贡献平均起来 (Shapley 1953)。p07 把它用于三位创始人:一位带来技术,一位带来销售渠道,一位带来资金与牌照。各个群体独自能创造的价值都是示意。

推导

# Chapter 4. p07: Shapley value of three founders. A = technology,
# B = sales channel, C = capital and licences. v(S) is the yearly value
# a coalition S could create on its own (illustrative): alone A 2, B 1,
# C 0; AB 8, AC 5, BC 3; ABC 12. The Shapley value averages each
# founder's marginal contribution over all orders of arrival. It is
# compared with an equal split and with a split by capital put in
# (A 10%, B 10%, C 80%).
from itertools import permutations

V = {"": 0, "A": 2, "B": 1, "C": 0, "AB": 8, "AC": 5, "BC": 3, "ABC": 12}

def key(s):
    return "".join(sorted(s))

def shapley(v, players="ABC"):
    out = {p: 0.0 for p in players}
    orders = list(permutations(players))
    for order in orders:
        have = ""
        for p in order:
            out[p] += v[key(have + p)] - v[key(have)]
            have += p
    return {p: x / len(orders) for p, x in out.items()}

if __name__ == "__main__":
    sh = shapley(V)
    total = V["ABC"]
    print("   founder   Shapley   share   equal   by capital")
    for p, cap in zip("ABC", (0.10, 0.10, 0.80)):
        print("   {:7}   {:7.3f}   {:5.3f}   {:5.3f}   {:10.3f}".format(
            p, sh[p], sh[p] / total, 1 / 3, cap))
    core = all(sum(sh[p] for p in s) >= V[key(s)] - 1e-9
               for s in ("A", "B", "C", "AB", "AC", "BC"))
    print("   Shapley split is in the core: {}".format(core))
    print("   AB alone could make {} > A + B under capital split {:.1f}"
          .format(V["AB"], 0.2 * total))
#    founder   Shapley   share   equal   by capital
#    A           5.667   0.472   0.333        0.100
#    B           4.167   0.347   0.333        0.100
#    C           2.167   0.181   0.333        0.800
#    Shapley split is in the core: True
#    AB alone could make 8 > A + B under capital split 2.4

沙普利值给技术创始人 12 单位中的 5.667(份额 0.472),销售创始人 4.167(0.347),资金创始人 2.167(0.181)。均分给每人 0.333,按出资分则给资金创始人 0.800。沙普利分配位于核心之中:没有任何群体离开会更好。按出资分配则不然:两位干活的创始人自己就能创造 8,却只分到公司价值的 2.4。资金创始人的钱有分量,但在这个博弈里,它的贡献小于另外两人的工作,而且通常可以作为投资来定价,而不是作为创始人股权。

观察漂移

如果有一位创始人悄悄撤出,同样的算术能否用作早期预警?p08 每个季度带噪声地测量每个群体的价值,并让创始人 C 从第 5 季度起撤出自己一半的贡献,使 C 的真实份额从 0.181 降到 0.111。

推导

# Chapter 4. p08: Watching contributions drift. Each quarter the value
# of every coalition is measured with noise (each v(S) times a
# lognormal factor, log-sd 0.25). From quarter 5 on, founder C quietly
# withdraws half of C's contribution: every v(S) containing C loses
# 1.5. C's Shapley share is recomputed each quarter; the alarm sounds
# when C's share averaged over the last k quarters falls 0.05 below its
# average in quarters 1 to 4. How long until it sounds, and how often
# it sounds by quarter 12 when nothing has changed? 4,000 runs, seed 93.
import random
from math import exp
from itertools import permutations

BASE = {"A": 2, "B": 1, "C": 0, "AB": 8, "AC": 5, "BC": 3, "ABC": 12}

def shapley_c(v):
    tot = 0.0
    for order in permutations("ABC"):
        have = ""
        for p in order:
            if p == "C":
                k1 = "".join(sorted(have + "C"))
                tot += v[k1] - (v["".join(sorted(have))] if have else 0)
            have += p
    return tot / 6 / v["ABC"]

def quarter(rng, q, change=True):
    v = {}
    for s, x in BASE.items():
        x = x - (1.5 if change and q >= 5 and "C" in s and x >= 1.5
                 else 0)
        v[s] = x * exp(rng.gauss(0, 0.25))
    return shapley_c(v)

def run(rng, k, change=True):
    shares = [quarter(rng, q, change) for q in range(1, 13)]
    base = sum(shares[:4]) / 4
    for q in range(4 + k, 13):
        if sum(shares[q - k:q]) / k < base - 0.05:
            return q, shares
    return None, shares

if __name__ == "__main__":
    rng = random.Random(93)
    print("   true C share: before {:.3f}, after {:.3f}".format(
        shapley_c({s: x for s, x in BASE.items()}),
        shapley_c({s: x - (1.5 if "C" in s and x >= 1.5 else 0)
                   for s, x in BASE.items()})))
    print("   window k   alarm by q8   by q12   median q   false alarm")
    for k in (1, 2, 4):
        qs = [run(rng, k)[0] for _ in range(4000)]
        hit = sorted(q for q in qs if q is not None)
        med = hit[len(hit) // 2] if hit else 0
        fa = sum(1 for _ in range(4000) if run(rng, k, False)[0]) / 4000
        print("   {:8d}   {:11.3f}   {:6.3f}   {:8d}   {:11.3f}".format(
            k, sum(1 for q in qs if q and q <= 8) / 4000,
            len(hit) / 4000, med, fa))
#    true C share: before 0.181, after 0.111
#    window k   alarm by q8   by q12   median q   false alarm
#           1         0.922    0.981          5         0.823
#           2         0.827    0.942          6         0.593
#           4         0.649    0.846          8         0.358

把每个季度与前四个季度比较的警报,在 C 真的撤出的情形中,到第 8 季度有 0.922 会响。但在什么都没有改变的情形中,到第 12 季度也有 0.823 会响。取两个季度平均,误报降到 0.593;取四个季度,降到 0.358,而到第 8 季度的发现率降到 0.649。实测贡献太嘈杂,不足以指控任何人;它适合用来讨论股权分配,而不能作为恶意的证据。

四次模拟中 C 的实测沙普利份额;虚线是第 5 季度前的真实份额,点线是之后的(p08)。

假说 · 尚未验证

作者的诠释(尚未验证): 以沙普利值等明确的贡献方法确定股权的创始团队,日后为股权分配发生的争议,少于按默认均分的团队。什么会证明它错: 在一项对创始团队的预先登记调查中,把团队人数与融资固定之后,确定股权的方法与日后的争议无关。

小结

沙普利值把「股权应当跟着贡献走」变成可以计算的分配,在例子里它与均分、按出资分都大不相同。但作为贡献变化的监测器,它太嘈杂:它的警报大多会是假的。

参考文献

Shapley, Lloyd S. 1953. “A Value for n-Person Games.” In Contributions to the Theory of Games, Volume II, edited by H. W. Kuhn and A. W. Tucker. Annals of Mathematics Studies 28. Princeton University Press.

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).