5D Library › 预测市场微观结构、布雷格曼信息投影与跨盘口无套利边界 · English
第一章 价格不是概率:信念分布、风险态度与冷热门偏差
一个流行的读法
预测市场的合约在事件发生时支付 1,否则支付 0。于是有一个流行的读法:价格 0.20 的合约,意味着市场认为事件有 20% 的可能。Wolfers 与 Zitzewitz 的综述让这个读法广为人知 (Wolfers and Zitzewitz 2004),它在不少场合也确实好用。但它是一个结论,需要前提。本章找出这些前提,并说明前提不成立时价格偏向哪里。
先把问题说清楚。市场里有许多交易者,各自对事件持有一个信念 。价格只有一个。把一个分布压缩为一个数,必须选定一个统计量:均值、中位数,或别的分位数。价格选的是哪一个,取决于交易者的预算、风险态度与信念分布的形状,而不取决于我们希望它是什么。
Manski 的结果:价格是分位数
Manski 用一个最简单的模型回答了这个问题 (Manski 2006):交易者风险中性、预算相同,信念高于价格的人把预算全部买 YES,低于价格的人全部买 NO。
推导
出清条件。 设信念的分布函数为 ,价格为 。每位 YES 买家用预算 1 买到 份,每位 NO 买家买到 份。YES 与 NO 的份数必须相等: 价格是这样一个点:信念高于它的交易者所占的比例恰好等于它自己。Manski 进一步证明,无论信念如何分布,信念均值都严格落在 之内——这是价格对均值所能说的全部。
推导
# Chapter 1. p01: Manski's result: a market price is not the mean belief.
# Risk-neutral traders hold equal budgets and a belief q that a binary
# event happens. A trader with q above the price spends the budget on YES
# contracts (each pays 1 if the event happens), one below it on NO.
# Clearing (YES bought = NO bought) gives F(price) = 1 - price: the price
# is the point where the SHARE of traders above it equals the price.
# Manski (2006): the mean belief lies strictly inside
# (price^2, 2*price - price^2), whatever the belief distribution.
# Section B gives budgets that rise with belief: then the price is the
# point where the WEALTH share above it equals the price.
import bisect
import random
N = 20000
def clear(beliefs, budgets=None):
pairs = sorted(zip(beliefs, budgets or [1.0] * len(beliefs)))
qs = [q for q, _ in pairs]
cum = [0.0]
for _, w in pairs:
cum.append(cum[-1] + w)
total = cum[-1]
lo, hi = 1e-9, 1 - 1e-9
for _ in range(60):
mid = (lo + hi) / 2
above = (total - cum[bisect.bisect_left(qs, mid)]) / total
lo, hi = (mid, hi) if above > mid else (lo, mid)
return (lo + hi) / 2
def draw(spec, rng):
if spec[0] == "beta":
return [rng.betavariate(spec[1], spec[2]) for _ in range(N)]
out = []
for _ in range(N):
a, b = spec[1] if rng.random() < spec[3] else spec[2]
out.append(rng.betavariate(a, b))
return out
if __name__ == "__main__":
rng = random.Random(59001)
pops = [("symmetric Beta(4,4)", ("beta", 4, 4)),
("pessimist Beta(2,8)", ("beta", 2, 8)),
("optimist Beta(8,2)", ("beta", 8, 2)),
("longshot Beta(1.5,13.5)", ("beta", 1.5, 13.5)),
("split 70/30 bimodal", ("mix", (2, 18), (18, 2), 0.7))]
print("A. equal budgets")
print(" population mean median price (p^2, 2p-p^2)")
kept = {}
for name, spec in pops:
q = draw(spec, rng)
kept[name] = q
p = clear(q)
mean, med = sum(q) / N, sorted(q)[N // 2]
ok = p * p < mean < 2 * p - p * p
print(" {:<26}{:5.3f} {:6.3f} {:5.3f} ({:5.3f}, {:5.3f}) {}"
.format(name, mean, med, p, p * p, 2 * p - p * p,
"inside" if ok else "OUTSIDE"))
print("\nB. Beta(1.5,13.5) beliefs; budget = q ** k (richer if bolder)")
q = kept["longshot Beta(1.5,13.5)"]
mean = sum(q) / N
for k in (0.0, 0.5, 1.0, 2.0):
w = [x ** k for x in q]
wmean = sum(a * b for a, b in zip(q, w)) / sum(w)
p = clear(q, w)
print(" k = {:3.1f} mean {:5.3f} wealth-weighted mean {:5.3f}"
" price {:5.3f}".format(k, mean, wmean, p))
print("\nReading: the price sits where the share (or wealth share) of")
print("traders above it equals it - a quantile, not the mean. With the")
print("same mean belief, skew and budgets move the price; only Manski's")
print("interval constrains the mean.")
# A. equal budgets
# population mean median price (p^2, 2p-p^2)
# symmetric Beta(4,4) 0.501 0.501 0.501 (0.251, 0.751) inside
# pessimist Beta(2,8) 0.201 0.181 0.267 (0.071, 0.462) inside
# optimist Beta(8,2) 0.800 0.822 0.735 (0.541, 0.930) inside
# longshot Beta(1.5,13.5) 0.100 0.083 0.167 (0.028, 0.306) inside
# split 70/30 bimodal 0.341 0.127 0.307 (0.094, 0.520) inside
#
# B. Beta(1.5,13.5) beliefs; budget = q ** k (richer if bolder)
# k = 0.0 mean 0.100 wealth-weighted mean 0.100 price 0.167
# k = 0.5 mean 0.100 wealth-weighted mean 0.128 price 0.193
# k = 1.0 mean 0.100 wealth-weighted mean 0.154 price 0.216
# k = 2.0 mean 0.100 wealth-weighted mean 0.202 price 0.257
#
# Reading: the price sits where the share (or wealth share) of
# traders above it equals it - a quantile, not the mean. With the
# same mean belief, skew and budgets move the price; only Manski's
# interval constrains the mean.A 段的五个人群说明了分位数与均值的差别。信念对称时,价格、均值、中位数重合于 0.501。信念偏斜时就不再如此:一个平均只信 0.100 的「冷门」人群,中位数 0.083,价格却是 0.167——比均值高出三分之二。原因在出清条件本身:价格低时,NO 方每份要付接近 1 的钱,YES 方每份只付很少,少数乐观者的预算就能买走大量 YES。同样的机制把乐观人群(均值 0.800)的价格压到 0.735。双峰人群的均值 0.341、中位数 0.127,价格 0.307,与两者都不同。五种情形的均值都落在 Manski 区间内,但区间很宽:冷门人群的区间是 (0.028, 0.306)。
B 段让预算随信念上升:越大胆的人越有钱。信念的分布不变,价格就从 0.167 一路升到 0.257。此时价格追随的是财富分布的分位数:财富位于价格之上的份额等于价格。
对数效用:唯一让价格等于均值的情形
Wolfers 与 Zitzewitz 给出了另一端的结论:若所有交易者都是对数效用,价格恰好等于以财富加权的平均信念 (Wolfers and Zitzewitz 2006)。Gjerstad 研究了风险厌恶如何使价格偏离这个均值 (Gjerstad 2005)。第二段程序用常相对风险厌恶(CRRA)效用逐一检验。
推导
CRRA 交易者的需求。 信念 、财富 、风险厌恶 的交易者持有 份 YES(负数为 NO)。最大化期望效用的一阶条件给出 价格使 。
推导
# Chapter 1. p02: Wolfers-Zitzewitz: with log utility the price IS the
# wealth-weighted mean belief; other risk attitudes move it away.
# Trader i has belief q, wealth w and CRRA risk aversion g. Holding n YES
# contracts bought at price p, terminal wealth is w + n(1 - p) if the
# event happens and w - n p if not. Maximising expected utility gives
# R = (q (1 - p) / ((1 - q) p)) ** (1 / g)
# n = w (R - 1) / ((1 - p) + R p)
# (n < 0 is a NO position). The price clears sum(n) = 0.
import math
import random
N = 4000
def demand(q, w, g, p):
r = (q * (1 - p) / ((1 - q) * p)) ** (1.0 / g)
return w * (r - 1) / ((1 - p) + r * p)
def clear(traders):
lo, hi = 1e-6, 1 - 1e-6
for _ in range(60):
mid = (lo + hi) / 2
excess = math.fsum(demand(q, w, g, mid) for q, w, g in traders)
lo, hi = (mid, hi) if excess > 0 else (lo, mid)
return (lo + hi) / 2
if __name__ == "__main__":
rng = random.Random(59002)
q = [rng.betavariate(1.5, 13.5) for _ in range(N)]
w = [math.exp(rng.gauss(0, 1)) for _ in range(N)]
mean = math.fsum(q) / N
wmean = math.fsum(a * b for a, b in zip(q, w)) / math.fsum(w)
med = sorted(q)[N // 2]
print("beliefs Beta(1.5,13.5): mean {:.4f}, wealth-weighted {:.4f},"
" median {:.4f}".format(mean, wmean, med))
print("\n risk aversion g price price - wtd mean")
cases = [("all g = 1 (log)", lambda x: 1.0),
("all g = 3", lambda x: 3.0),
("all g = 0.5 (risk-loving)", lambda x: 0.5),
("bold g=0.5, timid g=4", lambda x: 0.5 if x > med else 4.0),
("bold g=4, timid g=0.5", lambda x: 4.0 if x > med else 0.5)]
for name, gf in cases:
p = clear([(a, b, gf(a)) for a, b in zip(q, w)])
d = round(p - wmean, 4) + 0.0 # + 0.0 turns -0.0 into 0.0
print(" {:<26} {:6.4f} {:+.4f}".format(name, p, d))
print("\nReading: log utility returns the wealth-weighted mean exactly.")
print("Strong uniform risk aversion pulls the price toward the median")
print("belief, risk-loving toward the bold; when risk attitude goes with")
print("belief the price moves furthest, in either direction. A price")
print("reads as a probability only under an assumption about who trades")
print("and how they weigh risk.")
# beliefs Beta(1.5,13.5): mean 0.0984, wealth-weighted 0.0989, median 0.0819
#
# risk aversion g price price - wtd mean
# all g = 1 (log) 0.0989 +0.0000
# all g = 3 0.0823 -0.0167
# all g = 0.5 (risk-loving) 0.1171 +0.0182
# bold g=0.5, timid g=4 0.1444 +0.0454
# bold g=4, timid g=0.5 0.0560 -0.0429
#
# Reading: log utility returns the wealth-weighted mean exactly.
# Strong uniform risk aversion pulls the price toward the median
# belief, risk-loving toward the bold; when risk attitude goes with
# belief the price moves furthest, in either direction. A price
# reads as a probability only under an assumption about who trades
# and how they weigh risk.四千位交易者的信念与 p01 的冷门人群同分布,财富服从对数正态。对数效用下价格是 0.0989,与财富加权均值完全一致,这验证了程序本身。所有人风险厌恶系数为 3 时,价格降到 0.0823,几乎就是中位数 0.0819;所有人都偏好风险()时,价格升到 0.1171。最大的偏离出现在风险态度与信念相关时:大胆者同时不怕风险、谨慎者同时怕风险,价格是 0.1444,比财富加权均值高 0.0454;反过来则低 0.0429。
这说明「价格 = 概率」需要的不只是交易者整体无偏。它需要一个关于谁在交易、如何权衡风险的具体假设,而这个假设通常无法从价格本身检验。
冷热门偏差:不需要谁爱冒险
赛马投注中,冷门马的赔率长期过于慷慨,而热门马的赔率过于吝啬:同样一元,押冷门的平均回报低于押热门。这被称为冷热门偏差(favourite-longshot bias),最早由 Griffith 记录 (Griffith 1949),Ali 用赛马数据系统测量 (Ali 1977),Thaler 与 Ziemba 作了综述 (Thaler and Ziemba 1988)。常见的两种解释是:投注者爱好风险,或者投注者高估小概率 (Snowberg and Wolfers 2010)。
第三段程序检验第三种可能:什么偏好与认知偏差都没有,只有分歧。每位交易者看到的是真实概率加上无偏噪声。
推导
# Chapter 1. p03: The favourite-longshot bias from belief dispersion alone.
# 12,000 binary contracts. Each has a true probability p (uniform on the
# log-odds scale between -4 and +4). 200 traders per contract see p with
# unbiased noise on the log-odds scale (sd 1.0), so no one is biased on
# average and no one loves risk. Two clearing rules are compared:
# M Manski: risk-neutral, equal budgets (the quantile price of p01)
# L log utility, equal wealth (price = mean belief, as in p02)
# Contracts are grouped by price; each group's mean price is compared
# with its mean true probability and with simulated outcomes.
import bisect
import math
import random
K, T, SD = 12000, 200, 1.0
EDGES = (0.0, 0.05, 0.10, 0.20, 0.40, 0.60, 0.80, 0.90, 0.95, 1.0)
def sig(x):
return 1 / (1 + math.exp(-x))
def manski(qs):
qs = sorted(qs)
lo, hi = 1e-9, 1 - 1e-9
for _ in range(50):
mid = (lo + hi) / 2
above = 1 - bisect.bisect_left(qs, mid) / len(qs)
lo, hi = (mid, hi) if above > mid else (lo, mid)
return (lo + hi) / 2
def table(rows, label):
print("\n{} price band n mean price true p observed"
.format(label))
for lo, hi in zip(EDGES, EDGES[1:]):
g = [r for r in rows if lo <= r[0] < hi]
if not g:
continue
n = len(g)
mp = sum(r[0] for r in g) / n
tp = sum(r[1] for r in g) / n
ob = sum(r[2] for r in g) / n
print(" [{:.2f}, {:.2f}) {:5d} {:6.3f} {:6.3f} {:6.3f}"
.format(lo, hi, n, mp, tp, ob))
if __name__ == "__main__":
rng = random.Random(59003)
rows_m, rows_l = [], []
for _ in range(K):
x = rng.uniform(-4, 4)
p = sig(x)
qs = [sig(x + rng.gauss(0, SD)) for _ in range(T)]
y = 1 if rng.random() < p else 0
rows_m.append((manski(qs), p, y))
rows_l.append((sum(qs) / T, p, y))
table(rows_m, "M")
table(rows_l, "L")
print("\nReading: with unbiased but dispersed beliefs, both rules")
print("overprice longshots and underprice favourites - the bias needs")
print("neither risk-love nor misperception, only disagreement plus a")
print("price bounded by 0 and 1.")
#
# M price band n mean price true p observed
# [0.05, 0.10) 802 0.084 0.024 0.022
# [0.10, 0.20) 1905 0.145 0.058 0.065
# [0.20, 0.40) 2379 0.296 0.202 0.206
# [0.40, 0.60) 2039 0.500 0.499 0.502
# [0.60, 0.80) 2287 0.707 0.800 0.813
# [0.80, 0.90) 1798 0.854 0.942 0.946
# [0.90, 0.95) 790 0.914 0.976 0.977
#
# L price band n mean price true p observed
# [0.00, 0.05) 971 0.038 0.025 0.025
# [0.05, 0.10) 1212 0.072 0.050 0.049
# [0.10, 0.20) 1369 0.145 0.111 0.129
# [0.20, 0.40) 1799 0.294 0.260 0.260
# [0.40, 0.60) 1502 0.500 0.499 0.503
# [0.60, 0.80) 1708 0.705 0.738 0.752
# [0.80, 0.90) 1335 0.854 0.888 0.894
# [0.90, 0.95) 1178 0.927 0.950 0.954
# [0.95, 1.00) 926 0.962 0.975 0.977
#
# Reading: with unbiased but dispersed beliefs, both rules
# overprice longshots and underprice favourites - the bias needs
# neither risk-love nor misperception, only disagreement plus a
# price bounded by 0 and 1.两种出清规则都产生了冷热门偏差。Manski 规则下,价格在 0.05 到 0.10 之间的合约平均价 0.084,真实概率只有 0.024,模拟结果的实际发生率 0.022;价格 0.90 到 0.95 的合约平均价 0.914,真实概率 0.976。对数效用规则下偏差较小但方向相同:价格 0.038 的冷门真实概率 0.025,价格 0.962 的热门真实概率 0.975;中间一档(0.500 对 0.499)几乎没有偏差。
机制很简单。价格被夹在 0 与 1 之间,而噪声在对数几率上对称。真实概率很小时,噪声把一些人的信念推高很多、把另一些人的信念压低一点点——压低的那一侧已经贴近 0 了。对这些信念取平均(对数效用)或取 Manski 分位数,结果都被拉向 0.5。
注
这个结果说明了什么、没说明什么。 它说明冷热门偏差可以完全来自信念分散加上有界价格,因此在数据中看到偏差,并不能证明交易者爱好风险或高估小概率。它没有说明真实市场中偏差来自哪一种机制;Snowberg 与 Wolfers 用不同彩池投注的组合区分了风险偏好与概率误判两种解释 (Snowberg and Wolfers 2010),本书不重复那项工作。
假说 · 尚未验证
作者的假说(尚未验证): 在同一平台上,交易者之间信念分歧越大的合约(例如以同一事件不同来源的预测分散度衡量),冷门的价格高估越严重;分歧小的合约,冷热门偏差接近零。什么结果会证明它错:按分歧分组后,各组的偏差幅度相同,或分歧小的组偏差更大。
小结
预测市场的价格是信念分布的某个统计量,而这个统计量由交易规则决定。风险中性、预算相同时,价格是信念高于它的人所占比例等于它自己的那个分位数,信念均值只能被界定在 内;对数效用时,价格恰好是财富加权均值;其他风险态度,尤其是风险态度与信念相关时,价格可以朝任一方向偏离。只要信念有分歧、价格有上下界,冷门就会被高估、热门就会被低估,不需要任何人爱好风险。读价格之前,先问是谁在出价。
参考文献
Fan, Y. (2026). Prediction Market Microstructure, Bregman Information Projection and Cross-Book Arbitrage Bounds: Logarithmic Market Scoring Rules, Favorite-Longshot Calibration Bias, Lock-Up Opportunity Cost and Frank-Wolfe Combinatorial Pricing [预测市场微观结构、布雷格曼信息投影与跨盘口无套利边界] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090166 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).