5D Library › 多阶段最后通牒议价、应收账款衰减动力学与最优施压序列 · English

第一章 一串让步对对方透露了什么

阿克曼序列

一本流行的谈判书建议买方先定一个目标价,以它的 65% 开价,再升到 85%、95%,最后到 100%,并以一个精确的零头数字结尾 (Voss and Raz 2016)。它的说法是:递减的步子会让卖方相信买方已经接近底线。这是一个关于推断的说法,可以写成推断。设卖方把每一步看作以固定比例 kk 缩小上一次出价与买方真实底线 RR 之间的差距,外加一些噪声:

on+1−on=k(R−on)+ε.o_{n+1} – o_n = k\,(R – o_n) + \varepsilon.

没有噪声时,三次出价就能精确确定 RR:R=(o22−o1o3)/(2o2−o1−o3)R = (o_2^2 – o_1 o_3)/(2 o_2 – o_1 – o_3)。p01 给卖方一个宽的先验,并计算前三次出价之后的后验。

推导

# Chapter 1. p01: What a concession pattern tells the other side. A
# buyer with target price 100 makes four offers. Ackerman plan: 65, 85,
# 95, 100 (shrinking steps). Equal steps: 65, 77, 89, 100. The seller
# reads the steps with a simple model: each offer closes a fixed part
# k of the gap to the buyer's true limit R, plus noise (sd 2):
#   o(n+1) - o(n) = k (R - o(n)) + e.
# Prior: R uniform on 95 to 160, k uniform on 0.2 to 0.8 (grid). The
# seller's posterior for R after the first three offers: its mean, the
# chance that R is at most 105 and the chance that R is above 120.
from math import exp

def posterior(offers, sd=2.0):
    rs = [95 + i * 0.5 for i in range(131)]
    ks = [0.2 + j * 0.01 for j in range(61)]
    post = []
    for r in rs:
        w = 0.0
        for k in ks:
            ll = 0.0
            for a, b in zip(offers, offers[1:]):
                ll += -((b - a) - k * (r - a)) ** 2 / (2 * sd * sd)
            w += exp(ll)
        post.append(w)
    z = sum(post)
    return rs, [p / z for p in post]

if __name__ == "__main__":
    print("   plan             offers so far   E[R]    P(R <= 105)"
          "  P(R > 120)")
    for lab, offers in (("Ackerman", (65, 85, 95)),
                        ("equal steps", (65, 77, 89))):
        rs, p = posterior(offers)
        mean = sum(r * w for r, w in zip(rs, p))
        lo = sum(w for r, w in zip(rs, p) if r <= 105)
        hi = sum(w for r, w in zip(rs, p) if r > 120)
        print("   {:<15}  {:<14} {:6.1f}   {:10.3f}  {:10.3f}".format(
            lab, ", ".join(map(str, offers)), mean, lo, hi))
    o1, o2, o3 = 65, 85, 95
    print("   noise-free limit implied by 65, 85, 95: {:.1f}".format(
        (o2 * o2 - o1 * o3) / (2 * o2 - o1 - o3)))
#    plan             offers so far   E[R]    P(R <= 105)  P(R > 120)
#    Ackerman         65, 85, 95      110.9        0.389       0.185
#    equal steps      65, 77, 89      115.8        0.206       0.360
#    noise-free limit implied by 65, 85, 95: 105.0

出过 65、85、95 之后,卖方对底线的最佳估计是 110.9;他认为底线不超过 105 的机会是 0.389,超过 120 的是 0.185。等幅出价 65、77、89 之后,估计是 115.8,不超过 105 的只有 0.206,超过 120 的有 0.360。没有噪声时,65、85、95 意味着底线恰好是 105.0。递减的步子确实把卖方的信念推向买方的真实底线,但只推了一部分,而且只在卖方这样解读时才有效。知道这个方法、或者用别的方式解读步子的卖方,从中学不到任何东西。

递减出价与等幅出价之后,卖方对买方底线的后验(p01)。

精确的最后数字

实验发现,人们从精确的锚点出发调整得比从整数锚点少,仿佛精确意味着这个数字是算出来的 (Janiszewski and Uy 2008)。p02 把这一点写成似然比:假设处在真实底线的买方,比还有余地的买方更常报出精确的数字。

推导

# Chapter 1. p02: A precise final number. The buyer's last offer is
# either round (100) or precise (99.37, the same money to two decimal
# places). The seller asks whether it is truly final. Assumed
# likelihoods: buyers at their real limit state a precise number with
# chance 0.6 and a round one with 0.4; buyers with room left state a
# precise number with chance 0.2 and a round one with 0.8. Prior that
# the offer is final, 0.5 (after an Ackerman sequence, 0.7). The seller
# pushes for more unless the posterior that it is final exceeds 0.70;
# a push ends the talks with chance 0.15.
def final_belief(prior, precise):
    a = prior * (0.6 if precise else 0.4)
    b = (1 - prior) * (0.2 if precise else 0.8)
    return a / (a + b)

if __name__ == "__main__":
    print("   prior   last offer   P(final)   seller pushes?   deal chance")
    for prior in (0.5, 0.7):
        for precise in (False, True):
            q = final_belief(prior, precise)
            push = q <= 0.70
            print("   {:5.1f}   {:<10}   {:8.3f}   {:>14}   {:11.2f}".format(
                prior, "precise" if precise else "round", q,
                "yes" if push else "no", 0.85 if push else 1.00))
#    prior   last offer   P(final)   seller pushes?   deal chance
#      0.5   round           0.333              yes          0.85
#      0.5   precise         0.750               no          1.00
#      0.7   round           0.538              yes          0.85
#      0.7   precise         0.875               no          1.00

先验各半时,整数的最后报价让卖方认为「这是最后报价」的信念只有 0.333,卖方会要求更多,而这会以 0.15 的机会让谈判破裂;精确的数字把信念提高到 0.750,卖方就接受。在阿克曼序列已经把先验提高到 0.7 之后,两种信念分别是 0.538 与 0.875。精确数字只通过卖方对「谁会用精确数字」的假设起作用;如果虚张声势的买方也学会报精确数字,似然比与效果都会缩小到零。

假说 · 尚未验证

作者的诠释(尚未验证): 收到以精确数字结尾的递减让步序列的卖方,比收到以整数结尾、最终价格相同的等幅让步的卖方,对买方底线的估计更低,接受最后报价也更常见。什么会证明它错: 在一项随机分配出价序列、最终价格相同的预先登记谈判实验中,两种序列之间估计的底线与接受率都没有差别。

小结

递减的让步是关于底线的证据,精确的最后数字再多加一点,但两者都通过对方的推断起作用。在模型里它们的效果真实而有限,当对方知道这个方法时就消失了。

参考文献

Janiszewski, Chris, and Dan Uy. 2008. “Precision of the Anchor Influences the Amount of Adjustment.” Psychological Science 19 (2): 121 to 127.
Voss, Chris, and Tahl Raz. 2016. Never Split the Difference: Negotiating as If Your Life Depended on It. HarperBusiness.

Fan, Y. (2026). Multistage Ultimatum Bargaining, Accounts-Receivable Decay Kinetics and Optimal Escalation Sequences: Ackerman Concession Curves, Credible Threat Thresholds and Default Recovery Bounds [多阶段最后通牒议价、应收账款衰减动力学与最优施压序列] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090245 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).