5D Library › 结账顺手加购弹性、一键升单动力学与跨领域推荐矩阵 · English

第一章 结账加购的零摩擦套利

一个只需勾选的决定

买家打开收银台时,已经为主商品作了决定。此时在页面上再放一件相关的小商品,买家只需勾选一下,不必重新填写任何支付信息:加购的边际摩擦 κ\kappa 接近零。程序设主商品 RM 98,买家对加购品的估值为主商品价格乘以一个对数正态变量(中位数 0.45,对数标准差 0.8,示意参数);接受率为估值超过「价格 + 摩擦」的概率。

推导

# Chapter 1. p01: The order bump - optimal price ratio and take rate.
# At checkout the buyer has decided on the main item (price P = RM 98).
# The bump's value to a buyer is V = P x U, U lognormal (median 0.45,
# log-s.d. 0.8). Taking it costs the price p plus a friction kappa: near
# zero at checkout (one tick), larger later (re-entering card details).
# Take rate t(p) = P(V > p + kappa); bump revenue per order = p t(p).
from math import log, erf, sqrt

P = 98.0

def take(p, kappa=0.0, med=0.45, sd=0.8):
    x = (p + kappa) / P
    return 0.5 * (1 - erf((log(x) - log(med)) / sd / sqrt(2)))

def best(kappa=0.0, **kw):
    grid = [P * r / 1000 for r in range(50, 1500)]
    p = max(grid, key=lambda q: q * take(q, kappa, **kw))
    return p, take(p, kappa, **kw)

if __name__ == "__main__":
    for kappa in (0.0, 10.0, 25.0):
        p, t = best(kappa)
        print("   friction RM {:4.1f}: best bump RM {:6.2f} (ratio {:.3f}),"
              " take {:.3f}".format(kappa, p, p / P, t))
        print("     added to average order value: {:+.1%}".format(p * t / P))
    print("   ratio range over medians 0.30-0.60 and log-s.d. 0.6-1.0:")
    rs = []
    for med in (0.30, 0.45, 0.60):
        for sd in (0.6, 0.8, 1.0):
            p, t = best(0.0, med=med, sd=sd)
            rs.append((p / P, p * t / P))
    print("     ratio {:.3f} to {:.3f}; AOV lift {:+.1%} to {:+.1%}".format(
        min(r for r, _ in rs), max(r for r, _ in rs),
        min(a for _, a in rs), max(a for _, a in rs)))
#    friction RM  0.0: best bump RM  44.20 (ratio 0.451), take 0.499
#      added to average order value: +22.5%
#    friction RM 10.0: best bump RM  44.79 (ratio 0.457), take 0.393
#      added to average order value: +18.0%
#    friction RM 25.0: best bump RM  47.04 (ratio 0.480), take 0.270
#      added to average order value: +12.9%
#    ratio range over medians 0.30-0.60 and log-s.d. 0.6-1.0:
#      ratio 0.246 to 0.812; AOV lift +15.0% to +31.0%

摩擦为零时,使加购收入最大的价格是 RM 44.20,价格比 0.451,接受率 0.499,相当于把平均客单价提高 22.5%。同一件加购品若要在事后另行购买(摩擦 RM 10,例如重新填卡),最优价格几乎不变,但接受率降到 0.393,贡献降到 18.0%;摩擦 RM 25 时贡献只剩 12.9%。这就是「零摩擦套利」的含义:省下的不是价格,而是再做一次购买决定的成本。

最优价格比并不是一个普适常数。估值中位数在 0.30 到 0.60、对数标准差在 0.6 到 1.0 之间变化时,最优比从 0.246 到 0.812,对客单价的贡献从 +15.0% 到 +31.0%。它由买家对加购品的估值决定,需要在自己的流量上测出来(第六章)。

加购收入随价格比的变化:摩擦越小,曲线越高(p01)。

一个还是两个

推导

# Chapter 1. p02: One bump or two, and the cost to the main checkout.
# Each bump b has relative value U_b lognormal (medians 0.45 and 0.30,
# log-s.d. 0.8, correlation rho between the two log-values). Each bump
# shown lowers completion of the main checkout by a share delta (an extra
# decision on the page). Revenue per buyer who opens checkout:
#   (1 - delta n) (P + sum of bump prices x takes).
import random
from math import log, exp, sqrt

P = 98.0

def simulate(prices, meds, rho, delta, n=200000, seed=78002):
    rng = random.Random(seed)
    rev = takes0 = takes1 = 0.0
    for _ in range(n):
        z0 = rng.gauss(0, 1)
        z1 = rho * z0 + sqrt(1 - rho * rho) * rng.gauss(0, 1)
        vals = [P * meds[0] * exp(0.8 * z0), P * meds[1] * exp(0.8 * z1)]
        r = P
        for i, p in enumerate(prices):
            if vals[i] > p:
                r += p
                if i == 0:
                    takes0 += 1
                else:
                    takes1 += 1
        rev += r
    keep = 1 - delta * len(prices)
    return keep * rev / n, takes0 / n, takes1 / n

if __name__ == "__main__":
    print("   bumps  rho   delta  takes         revenue/opened  vs none")
    base = P
    for prices, rho, delta in (((44.0,), 0.0, 0.01),
                               ((44.0, 29.0), 0.0, 0.01),
                               ((44.0, 29.0), 0.5, 0.01),
                               ((44.0, 29.0), 0.5, 0.03)):
        r, t0, t1 = simulate(prices, (0.45, 0.30), rho, delta)
        print("   {:5d}  {:.1f}  {:5.2f}  {:.3f} {:.3f}   {:14.2f}  {:+.1%}"
              .format(len(prices), rho, delta, t0,
                      t1 if len(prices) > 1 else 0, r, r / base - 1))
#    bumps  rho   delta  takes         revenue/opened  vs none
#        1  0.0   0.01  0.501 0.000           118.85  +21.3%
#        2  0.0   0.01  0.501 0.505           132.01  +34.7%
#        2  0.5   0.01  0.501 0.506           132.04  +34.7%
#        2  0.5   0.03  0.501 0.506           126.65  +29.2%

程序在结账页放一个(RM 44)或两个(RM 44 与 RM 29)加购,并假设每多放一个,主商品的完成率下降 δ\delta。δ=0.01\delta = 0.01 时,一个加购使每位打开结账者的收入从 RM 98 提高到 RM 118.85(+21.3%),两个加购 RM 132.01(+34.7%);若每多一个选项使完成率下降 0.03,两个加购只有 +29.2%。两个加购估值的相关性在这个模型中不改变收入,因为两件商品分别定价、分别接受。

注

加购必须是买家主动勾选的。 预先勾选的加购会让一部分买家在不知情的情况下付款,属于有记录的暗模式,也会带来退款与拒付(第 77 本第五章)。本章的模型只适用于默认不勾选、价格清楚可见的加购。

小结

结账加购的优势来自零摩擦:示意模型中最优价格比 0.451、接受率 0.499,客单价 +22.5%;最优比随估值分布在 0.246 到 0.812 之间变化。第二个加购再加十几个百分点,但每个额外选项都要从主商品的完成率中付出代价。

Fan, Y. (2026). Order-Bump Friction Elasticity, One-Click Upsell Kinetics and Cross-Domain Recommendation Matrices: Zero-Reentry Impulse Attachment, Conditional Mutual Information and Cart-Value Multipliers [结账顺手加购弹性、一键升单动力学与跨领域推荐矩阵] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090207 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).