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

第三章 产品线自蚕食:嵌套 Logit 与转移分流率

十个 SKU 的需求模型

一家店铺卖十个 SKU:一款旗舰品 F,四款与它相近的副线 A1 至 A4(更便宜、毛利更低),以及五款低价品 B1 至 B5。买家也可以去别家买或者不买,这是「外部选项」。嵌套 Logit 模型把彼此相近的商品放在同一个巢里 (McFadden 1978; Berry 1994):巢内的商品彼此替代更强,巢参数 λg\lambda_g 越小,巢内替代越强。商品 jj 的份额是

sj=eδj/λgDg⋅Dgλgeu0+∑hDhλh,Dg=∑k∈geδk/λg,s_j = \frac{e^{\delta_j/\lambda_g}}{D_g}\cdot\frac{D_g^{\lambda_g}}{e^{u_0} + \sum_h D_h^{\lambda_h}},\qquad D_g = \sum_{k \in g} e^{\delta_k/\lambda_g},

其中 δj\delta_j 是质量减去价格乘以价格敏感度,u0u_0 是外部选项的效用。

衡量自蚕食最直接的量是转移分流率:商品 jj 下架之后,它原来的买家有多少比例转去买商品 kk (Farrell and Shapiro 2010)。

推导

# Chapter 3. p05: Nested logit demand for ten SKUs and diversion ratios.
# Nest A holds the flagship and four sub-line variants (close substitutes,
# lambda 0.6); nest B holds five budget items (lambda 0.85). Utility is
# q_j - 0.02 * price; the outside option (rival shops, or buying nothing)
# has utility u0 = 2.0.
# Diversion j -> k when j is delisted: (s_k without j - s_k) / s_j.
from math import exp

SKUS = (  # name, nest, price RM, unit margin RM, quality
    ("F", "A", 299, 120, 6.9), ("A1", "A", 259, 90, 5.7),
    ("A2", "A", 229, 60, 5.25), ("A3", "A", 199, 40, 4.5),
    ("A4", "A", 189, 32, 4.2), ("B1", "B", 129, 40, 2.2),
    ("B2", "B", 119, 34, 1.9), ("B3", "B", 109, 28, 1.6),
    ("B4", "B", 99, 22, 1.3), ("B5", "B", 89, 16, 0.9))
LAM = {"A": 0.6, "B": 0.85}
U0 = 2.0
SHOPPERS = 10000

def shares(active, lam=LAM, skus=SKUS, u0=U0):
    d = {s[0]: s[4] - 0.02 * s[2] for s in skus if s[0] in active}
    nest = {s[0]: s[1] for s in skus}
    D = {}
    for j, v in d.items():
        D[nest[j]] = D.get(nest[j], 0.0) + exp(v / lam[nest[j]])
    top = exp(u0) + sum(Dg ** lam[g] for g, Dg in D.items())
    out = {}
    for j, v in d.items():
        g = nest[j]
        out[j] = exp(v / lam[g]) / D[g] * D[g] ** lam[g] / top
    out["none"] = exp(u0) / top
    return out

def diversion(j, active, lam=LAM, skus=SKUS, u0=U0):
    s = shares(active, lam, skus, u0)
    s2 = shares([a for a in active if a != j], lam, skus, u0)
    return {k: (s2[k] - s[k]) / s[j] for k in s2}

ALL = [s[0] for s in SKUS]

if __name__ == "__main__":
    s = shares(ALL)
    print("   SKU  share  units/month")
    for name in ALL + ["none"]:
        print("   {:<4} {:.3f}  {:7.0f}".format(name, s[name],
                                              SHOPPERS * s[name]))
    for j in ("A4", "A3", "B5"):
        dv = diversion(j, ALL)
        top = sorted(dv.items(), key=lambda kv: -kv[1])[:3]
        print("   delist {}: diversion ".format(j) + ", ".join(
            "{} {:.3f}".format(k, v) for k, v in top))
#    SKU  share  units/month
#    F    0.109     1094
#    A1   0.056      562
#    A2   0.072      721
#    A3   0.056      562
#    A4   0.048      475
#    B1   0.038      383
#    B2   0.034      340
#    B3   0.030      302
#    B4   0.027      269
#    B5   0.021      213
#    none 0.508     5079
#    delist A4: diversion none 0.323, F 0.216, A2 0.143
#    delist A3: diversion none 0.327, F 0.221, A2 0.146
#    delist B5: diversion none 0.445, F 0.096, B1 0.076

在示意参数下,每月 10,000 名进店买家中,旗舰品的份额是 0.109(1,094 件),外部选项是 0.508。下架副线 A4,它的买家中有 0.323 去了别家或不买,0.216 转买旗舰品,0.143 转买 A2。下架 A3,转向旗舰品的比例是 0.221。下架低价品 B5,0.445 流向外部,只有 0.096 转向旗舰品,0.076 转向 B1。

砍掉尾部三成

经验法则说:按贡献排序,砍掉尾部 30% 的 SKU,利润往往不降反升。每个在售 SKU 每月要花 RM 2,500 维持(库存、上架、拍摄与客服)。按「销量乘以单件毛利减去维持成本」排序,尾部三个是 B5、B4、B3。

推导

# Chapter 3. p06: Pareto pruning - delist the 30% of SKUs (3 of 10) with
# the lowest contribution (units x margin - listing cost) and recompute
# profit with the demand of p05. Each listed SKU costs RM 2,500 a month
# (inventory, listings, photos, support). The gain is split into its
# parts; then the flagship is capped at its current output (no spare
# capacity), so demand diverted to it beyond the cap is lost.
import importlib.util

def load(name):
    spec = importlib.util.spec_from_file_location(name, name + ".py")
    m = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(m)
    return m

p05 = load("p05_nested_logit")
FIXED = 2500.0
MARGIN = {s[0]: s[3] for s in p05.SKUS}
PRICE = {s[0]: s[2] for s in p05.SKUS}
N = p05.SHOPPERS

def units(active, cap_f=None):
    s = p05.shares(active)
    u = {j: N * s[j] for j in active}
    if cap_f is not None and "F" in u:
        u["F"] = min(u["F"], cap_f)
    return u

def profit(active, cap_f=None):
    u = units(active, cap_f)
    return sum(u[j] * MARGIN[j] for j in u) - FIXED * len(active)

def tail(active, k):
    u = units(active)
    return sorted(active, key=lambda j: u[j] * MARGIN[j] - FIXED)[:k]

if __name__ == "__main__":
    allv = list(p05.ALL)
    cut = tail(allv, 3)
    kept = [j for j in allv if j not in cut]
    u0, u1 = units(allv), units(kept)
    p0, p1 = profit(allv), profit(kept)
    print("   delist {} (lowest contribution)".format(", ".join(cut)))
    print("   units {:,.0f} -> {:,.0f}; revenue RM {:,.0f} -> {:,.0f}".format(
        sum(u0.values()), sum(u1.values()),
        sum(u0[j] * PRICE[j] for j in u0),
        sum(u1[j] * PRICE[j] for j in u1)))
    print("   profit RM {:,.0f} -> {:,.0f} ({:+.1%})".format(
        p0, p1, p1 / p0 - 1))
    lost = sum(u0[j] * MARGIN[j] for j in cut)
    to_f = (u1["F"] - u0["F"]) * MARGIN["F"]
    other = sum((u1[j] - u0[j]) * MARGIN[j] for j in kept if j != "F")
    print("   parts: listing costs saved RM {:,.0f}".format(3 * FIXED))
    print("          margin lost on delisted SKUs RM {:,.0f}".format(-lost))
    print("          recaptured by the flagship RM {:,.0f}".format(to_f))
    print("          recaptured by other SKUs RM {:,.0f}".format(other))
    cap = u0["F"]
    q0, q1 = profit(allv, cap), profit(kept, cap)
    print("   flagship capped at {:,.0f} units: RM {:,.0f} -> {:,.0f}"
          " ({:+.1%})".format(cap, q0, q1, q1 / q0 - 1))
#    delist B5, B4, B3 (lowest contribution)
#    units 4,921 -> 4,539; revenue RM 1,007,776 -> 1,010,436
#    profit RM 282,460 -> 297,334 (+5.3%)
#    parts: listing costs saved RM 7,500
#           margin lost on delisted SKUs RM -17,786
#           recaptured by the flagship RM 9,877
#           recaptured by other SKUs RM 15,282
#    flagship capped at 1,094 units: RM 282,460 -> 287,456 (+1.8%)

下架这三款之后,销量从 4,921 件降到 4,539 件,营业收入从 RM 1,007,776 升到 RM 1,010,436,利润从 RM 282,460 升到 RM 297,334(+5.3%)。把这笔变化拆开:省下维持成本 RM 7,500;失去这三款的毛利 RM 17,786;旗舰品接回 RM 9,877;其他在售商品接回 RM 15,282。利润上升主要来自转移分流:一部分买家转去买毛利更高的商品。

注

分流要有地方接。 如果旗舰品已经满负荷,多出来的需求接不住,下架的收益就小得多:把旗舰品的供应封顶在现在的 1,094 件,同样的下架只让利润从 RM 282,460 升到 RM 287,456(+1.8%)。这与第一章是同一个问题:分流过去的需求要经过瓶颈。

小结

在示意的嵌套 Logit 模型中,砍掉贡献最低的三成 SKU 使利润上升 5.3%,原因是一部分买家转去买毛利更高的商品,再加上省下的维持成本。这个结论依赖转移分流率与接收商品的产能;下一章检查它在其他情形下是否成立。

参考文献

Berry, Steven T. 1994. “Estimating Discrete-Choice Models of Product Differentiation.” The RAND Journal of Economics 25 (2): 242–62.
Farrell, Joseph, and Carl Shapiro. 2010. “Antitrust Evaluation of Horizontal Mergers: An Economic Alternative to Market Definition.” The B.E. Journal of Theoretical Economics 10 (1).
McFadden, Daniel. 1978. “Modelling the Choice of Residential Location.” In Spatial Interaction Theory and Planning Models, edited by Anders Karlqvist, Lars Lundqvist, Folke Snickars, and Jörgen W. Weibull. North-Holland.

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).