5D Library › 企业瓶颈诊断、产品线自蚕食审计与自动化改造回报边界 · English

第六章 按效计费:阶梯分成、风险偏好与基线测量

线性、阶梯还是固定费

按效计费(value-based pricing)的合约有三种常见形式:线性分成;超过门槛之后才分成的阶梯分成;与效果无关的固定费。三者对顾问的意义不同:期望报酬、报酬的波动、以及边际激励(每多创造一单位预期增量,报酬多多少)。

推导

# Chapter 6. p11: Linear share, stepped share or a flat fee, when the
# measured monthly gain M is normal with mean 86.3 and sd 40 (RM 1k).
#   linear : 8.1 + 0.214 * M
#   stepped: 8.1 + 0.35 * max(M - 60, 0)
#   flat   : 26.6 (same expected pay as the linear contract)
# For a normal X, E[max(X - T, 0)] = (mu - T) Phi(d) + sd phi(d),
# d = (mu - T) / sd. Reported: expected pay, its sd, the incentive (pay
# per extra RM 1 of mean gain), and what a riskier plan (sd 60) is worth.
from math import erf, exp, pi, sqrt

def Phi(x):
    return 0.5 * (1 + erf(x / sqrt(2)))

def phi(x):
    return exp(-x * x / 2) / sqrt(2 * pi)

def call(mu, sd, t):
    d = (mu - t) / sd
    return (mu - t) * Phi(d) + sd * phi(d)

def call2(mu, sd, t):
    """E[max(X - T, 0)^2] for the variance of the stepped pay."""
    d = (mu - t) / sd
    m = mu - t
    return (m * m + sd * sd) * Phi(d) + m * sd * phi(d)

def stats(kind, mu, sd):
    if kind == "linear":
        return 8.1 + 0.214 * mu, 0.214 * sd, 0.214
    if kind == "stepped":
        e1 = call(mu, sd, 60.0)
        v = 0.35 ** 2 * (call2(mu, sd, 60.0) - e1 * e1)
        return 8.1 + 0.35 * e1, sqrt(v), 0.35 * Phi((mu - 60.0) / sd)
    return 26.6, 0.0, 0.0

if __name__ == "__main__":
    print("   contract  E[pay]  sd(pay)  incentive  E[pay] if sd = 60")
    for k in ("linear", "stepped", "flat"):
        e, s, slope = stats(k, 86.3, 40.0)
        e60 = stats(k, 86.3, 60.0)[0]
        print("   {:<8}  {:6.1f}  {:7.1f}  {:9.3f}  {:6.1f} ({:+.1f})".format(
            k, e, s, slope, e60, e60 - e))
    for mu in (40.0, 60.0, 120.0):
        print("   stepped incentive at mean gain {:.0f}: {:.3f}".format(
            mu, stats("stepped", mu, 40.0)[2]))
#    contract  E[pay]  sd(pay)  incentive  E[pay] if sd = 60
#    linear      26.6      8.6      0.214    26.6 (+0.0)
#    stepped     19.5     11.0      0.261    21.9 (+2.4)
#    flat        26.6      0.0      0.000    26.6 (+0.0)
#    stepped incentive at mean gain 40: 0.108
#    stepped incentive at mean gain 60: 0.175
#    stepped incentive at mean gain 120: 0.327

在第五章的参数下(月度增量均值 86.3、标准差 40,单位千令吉):线性合约 8.1+0.214M8.1 + 0.214M 的期望报酬 26.6、标准差 8.6、边际激励 0.214;阶梯合约 8.1+0.35max⁡(M−60,0)8.1 + 0.35\max(M-60, 0) 的期望报酬只有 19.5,标准差却有 11.0,边际激励 0.261;固定费 26.6 没有波动,也没有激励。

阶梯合约有两个副作用。第一,它是凸的,所以奖励波动:如果顾问选择一个风险更大的方案,使增量的标准差从 40 升到 60,而均值不变,阶梯合约的期望报酬从 19.5 升到 21.9(+2.4),线性合约不变。第二,它的激励随项目进度而变:均值远低于门槛时(40),边际激励只有 0.108;均值 60 时 0.175;均值 120 时 0.327。项目一旦看来达不到门槛,顾问的激励就近乎消失。

注

凸的合约奖励冒险。 若客户不希望顾问为了跨过门槛而选择高风险的方案(例如一次性大幅砍掉 SKU),线性分成比阶梯分成更稳妥。

增量怎么量:基线、对照组与观测月数

分成的前提是能把增量量出来。月度毛利有噪声;比较改造前后各 mm 个月,增量估计的标准误是 σeff2/m\sigma_{\text{eff}}\sqrt{2/m}。如果找到一个对照(另一条产品线、另一家门店),它的月度噪声与客户的相关系数为 ρ\rho,用双重差分可以把有效噪声降到 σ2(1−ρ)\sigma\sqrt{2(1-\rho)}。

推导

# Chapter 6. p12: How long must the gain be measured before it can be
# paid on? Monthly gross profit has noise sd 40 (RM 1k). Comparing m
# months after with m months before gives SE = sd_eff * sqrt(2 / m); a
# comparison shop or product line whose monthly noise correlates rho with
# the client's (difference in differences) cuts sd_eff to
# sd * sqrt(2 (1 - rho)). Power 0.80 at two-sided alpha 0.05 needs
# effect / SE >= 1.96 + 0.84.
from math import ceil, sqrt

SD, Z = 40.0, 1.96 + 0.84

def sd_eff(rho):
    return SD if rho is None else SD * sqrt(2 * (1 - rho))

def months(effect, rho):
    return ceil(2 * (Z * sd_eff(rho) / effect) ** 2)

def mde(m, rho):
    return Z * sd_eff(rho) * sqrt(2 / m)

if __name__ == "__main__":
    print("   comparison        months each side to detect RM 30k / 60k;"
          " MDE at 6")
    for label, rho in (("before-after only", None), ("control rho 0.5", 0.5),
                       ("control rho 0.8", 0.8), ("control rho 0.9", 0.9)):
        print("   {:<17} {:5d} / {:3d}   RM {:5.1f}k".format(
            label, months(30.0, rho), months(60.0, rho), mde(6, rho)))
#    comparison        months each side to detect RM 30k / 60k; MDE at 6
#    before-after only    28 /   7   RM  64.7k
#    control rho 0.5      28 /   7   RM  64.7k
#    control rho 0.8      12 /   3   RM  40.9k
#    control rho 0.9       6 /   2   RM  28.9k

月度噪声标准差 40 时,只做前后比较,要在 80% 的把握下检出每月 RM 30k 的增量,前后各需 28 个月;检出 RM 60k 需 7 个月;前后各 6 个月时的最小可检出增量是 RM 64.7k。对照的相关系数只有 0.5 时毫无帮助(结果与只做前后比较相同);相关系数 0.8 时,检出 RM 30k 只需 12 个月,6 个月的最小可检出增量是 RM 40.9k;相关系数 0.9 时只需 6 个月,最小可检出增量 RM 28.9k。

假说 · 尚未验证

作者的假说(尚未验证): 以对照产品线或对照门店做双重差分的分成合约,比只做前后比较的合约,结算争议更少。什么会证明它错: 两类合约中客户对结算金额提出异议的比例没有差别。

小结

线性分成的激励稳定,阶梯分成在门槛附近激励最强、在门槛以下几乎为零,并且奖励冒险。分成之前要先把增量量准:只做前后比较往往需要两年以上,一个相关系数高于 0.5 的对照组才能真正缩短测量期。

Fan, Y. (2026). Enterprise Bottleneck Diagnostics, SKU Cannibalization Audits and Automation Return Bounds: Theory-of-Constraints Queueing Networks, Margin Dilution Matrices and Value-Based Advisory Pricing [企业瓶颈诊断、产品线自蚕食审计与自动化改造回报边界] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090217 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).