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
第三章 资金锁仓:为什么几乎确定的合约停在 0.92 到 0.96
一个看似白送的钱
距离结算还有半年,一份几乎确定会兑现的合约报价 0.95。如果它的真实概率是 0.99,买入就有 4% 的期望收益;这似乎是白送的钱,应当立刻被套利者买到 0.99。可是这类合约常常在 0.92 到 0.96 之间停留数月,直到结算前几周才向 1 收敛。本章说明这不是市场失灵,而是资金成本的直接结果。
关键在于:合约在结算之前不付利息。买入价格为 的合约,就是把 锁在里面,直到结算那天。资金另有用途——无风险利率,或者交易者在别处能赚到的回报。Shleifer 与 Vishny 称这类约束为「套利的极限」:套利需要资本,而资本有成本 (Shleifer and Vishny 1997)。
无套利带
推导
锁仓折现。 合约在 年后结算,以概率 支付 1。交易者的资金在别处每年可得 ,所以一个头寸必须提供 的总回报才值得持有。以价格 买 YES,期望回报为 ;以 买 NO,期望回报为 。两方都不愿交易的区间是 在这个区间之内,价格由确实想交易的人决定:对冲者、支持者、买冷门的人。
推导
# Chapter 3. p06: Capital lock-up opens a no-arbitrage band around p.
# A contract pays 1 at resolution, tau years away, with probability p.
# Money in it earns nothing until then; the trader's alternative earns
# rho a year, so a position must return g = (1 + rho) ** tau to be worth
# holding. Buying YES at price x pays p / x on average; buying NO at
# 1 - x pays (1 - p) / (1 - x). Neither side wants to trade while
# p / g <= x <= 1 - (1 - p) / g.
# Inside the band the price is set by whoever does want to trade: hedgers,
# fans, longshot buyers. rho = 0.045 is a risk-free rate; rho = 0.145
# adds a 10 % a year return the trader could earn elsewhere.
TAUS = ((1 / 52, "1 week"), (1 / 12, "1 month"), (0.25, "3 months"),
(0.5, "6 months"), (1.0, "1 year"), (2.0, "2 years"))
def band(p, tau, rho):
g = (1 + rho) ** tau
return p / g, 1 - (1 - p) / g
if __name__ == "__main__":
for p in (0.99, 0.05):
print("true p = {:.2f}".format(p))
print(" horizon rho = 0.045 rho = 0.145")
for tau, name in TAUS:
a = band(p, tau, 0.045)
c = band(p, tau, 0.145)
print(" {:<9} [{:.4f}, {:.4f}] [{:.4f}, {:.4f}]"
.format(name, a[0], a[1], c[0], c[1]))
print()
p, tau = 0.99, 0.5
for x in (0.92, 0.95, 0.96):
r = (p / x) ** (1 / tau) - 1
print("buy YES at {:.2f}, 6 months, p = 0.99: expected {:+.1%}"
" a year".format(x, r))
print("\nReading: a near-certain contract a year out can trade anywhere")
print("from 0.86 to 0.99 without offering anyone a return above 14.5 %;")
print("a 5 % longshot can trade up to 0.17. The band, not the")
print("probability, is what arbitrage enforces, and it closes only as")
print("resolution nears.")
# true p = 0.99
# horizon rho = 0.045 rho = 0.145
# 1 week [0.9892, 0.9900] [0.9874, 0.9900]
# 1 month [0.9864, 0.9900] [0.9789, 0.9901]
# 3 months [0.9792, 0.9901] [0.9570, 0.9903]
# 6 months [0.9684, 0.9902] [0.9252, 0.9907]
# 1 year [0.9474, 0.9904] [0.8646, 0.9913]
# 2 years [0.9066, 0.9908] [0.7551, 0.9924]
#
# true p = 0.05
# horizon rho = 0.045 rho = 0.145
# 1 week [0.0500, 0.0508] [0.0499, 0.0525]
# 1 month [0.0498, 0.0535] [0.0494, 0.0607]
# 3 months [0.0495, 0.0604] [0.0483, 0.0816]
# 6 months [0.0489, 0.0707] [0.0467, 0.1122]
# 1 year [0.0478, 0.0909] [0.0437, 0.1703]
# 2 years [0.0458, 0.1301] [0.0381, 0.2754]
#
# buy YES at 0.92, 6 months, p = 0.99: expected +15.8% a year
# buy YES at 0.95, 6 months, p = 0.99: expected +8.6% a year
# buy YES at 0.96, 6 months, p = 0.99: expected +6.3% a year
#
# Reading: a near-certain contract a year out can trade anywhere
# from 0.86 to 0.99 without offering anyone a return above 14.5 %;
# a 5 % longshot can trade up to 0.17. The band, not the
# probability, is what arbitrage enforces, and it closes only as
# resolution nears.第一张表给出真实概率 0.99 的合约。只按无风险利率()计,一年期合约可以在 0.9474 到 0.9904 之间任意定价而不给任何人超过 4.5% 的年回报。若交易者在别处还能多赚 10%(),区间的下沿降到 0.8646;半年期为 0.9252,三个月期为 0.9570。0.92 到 0.96 正落在三个月到半年期合约的下沿附近。
第二张表把同样的计算用于真实概率只有 0.05 的冷门。一年期、 时,区间上沿是 0.1703:一份 5% 的冷门可以卖到 0.17 而不被套利,因为纠正它需要买 NO——以 的价格锁一年资金,只为了赚几个百分点。
最后三行把「白送的钱」折成年化回报。半年期、真实概率 0.99 的合约,以 0.95 买入,年化 +8.6%;以 0.96 买入,+6.3%;只有跌到 0.92,年化才达到 +15.8%。对一位在别处能赚 14.5% 的交易者,0.95 并不便宜。
注
这里没有计入的东西。 区间只计入了资金的时间成本。平台风险(交易所倒闭、结算争议、预言机出错)、手续费与提现成本会让区间更宽;能用合约作保证金、或持有期间可获利息的平台会让它更窄。 因人而异,区间的位置也因人而异:一个只能拿到无风险利率的交易者,会愿意在 0.9474 买入一年期合约,而一位能赚 14.5% 的交易者不会。
冷热门偏差为什么能存活
第一章说明,信念分散足以产生冷热门偏差。如果资本没有成本,知道真实概率的套利者会把偏差纠正干净。第七段程序把两者放在一起:价格由人群的平均信念决定,但当它超出无套利带时,有知情资本把它推回带的边缘。
推导
# Chapter 3. p07: Why long-dated prices lean toward 0.5.
# Each contract has a true p (uniform on log-odds in [-4, 4]). The crowd
# price is the mean of 200 noisy beliefs (log-odds noise sd 1.0, the L
# rule of p03). Informed arbitrage capital knows p and requires
# rho = 0.145 a year, so it pushes the crowd price only back to the edge
# of the no-arbitrage band of p06; inside the band the crowd price stands.
# With rho = 0 the band is a point and prices equal p. Calibration error
# (mean price minus mean true p) is reported by price band and horizon.
import math
import random
K, T, SD, RHO = 6000, 200, 1.0, 0.145
EDGES = (0.0, 0.1, 0.3, 0.7, 0.9, 1.0)
def sig(x):
return 1 / (1 + math.exp(-x))
def priced(rng, tau, rho):
rows = []
for _ in range(K):
x = rng.uniform(-4, 4)
p = sig(x)
crowd = sum(sig(x + rng.gauss(0, SD)) for _ in range(T)) / T
g = (1 + rho) ** tau
lo, hi = p / g, 1 - (1 - p) / g
rows.append((min(max(crowd, lo), hi), p))
return rows
if __name__ == "__main__":
rng = random.Random(59007)
heads = "".join(" [{:.1f},{:.1f})".format(a, b)
for a, b in zip(EDGES, EDGES[1:]))
print("mean price - mean true p, by price band")
print(" horizon rho " + heads)
for tau, name in ((1 / 12, "1 month"), (0.5, "6 months"),
(1.0, "1 year"), (2.0, "2 years")):
for rho in (0.0, RHO):
rows = priced(rng, tau, rho)
cells = []
for a, b in zip(EDGES, EDGES[1:]):
g = [r for r in rows if a <= r[0] < b]
d = round(sum(r[0] - r[1] for r in g) / len(g), 4) + 0.0
cells.append(" {:+9.4f}".format(d))
print(" {:<8} {:5.3f}".format(name, rho) + "".join(cells))
print("\nReading: with costless capital the crowd's bias is arbitraged")
print("away; with a 14.5 % hurdle it survives - longshots too dear,")
print("favourites too cheap - growing with the horizon until the band")
print("is wide enough to hold the whole crowd bias (by six months here).")
print("The same price means a different probability a month out.")
# mean price - mean true p, by price band
# horizon rho [0.0,0.1) [0.1,0.3) [0.3,0.7) [0.7,0.9) [0.9,1.0)
# 1 month 0.000 +0.0000 +0.0000 +0.0000 +0.0000 +0.0000
# 1 month 0.145 +0.0105 +0.0092 -0.0001 -0.0092 -0.0106
# 6 months 0.000 +0.0000 +0.0000 +0.0000 +0.0000 +0.0000
# 6 months 0.145 +0.0183 +0.0358 +0.0002 -0.0355 -0.0181
# 1 year 0.000 +0.0000 +0.0000 +0.0000 +0.0000 +0.0000
# 1 year 0.145 +0.0179 +0.0354 -0.0005 -0.0355 -0.0178
# 2 years 0.000 +0.0000 +0.0000 +0.0000 +0.0000 +0.0000
# 2 years 0.145 +0.0183 +0.0351 +0.0004 -0.0357 -0.0181
#
# Reading: with costless capital the crowd's bias is arbitraged
# away; with a 14.5 % hurdle it survives - longshots too dear,
# favourites too cheap - growing with the horizon until the band
# is wide enough to hold the whole crowd bias (by six months here).
# The same price means a different probability a month out.资本没有成本()时,带缩成一个点,所有偏差都被消除,每一格都是 0。资本要求 14.5% 时,偏差存活下来:一个月期的冷门平均高估 +0.0105、热门低估 −0.0106;半年期,价格在 0.1 到 0.3 之间的合约高估 +0.0358,0.7 到 0.9 之间的低估 −0.0355。偏差随期限增长,直到带宽足以容纳人群全部的偏差(这里在半年时已经达到),此后一年期与两年期都与半年期相近。
这正是 Page 与 Clemen 在真实预测市场中发现的模式:距离结算越远,价格越向 0.5 倾斜,越近越接近校准 (Page and Clemen 2013)。本章的模型给出一个机制:冷门的高估与热门的低估都在无套利带内,而带宽随期限增长。
假说 · 尚未验证
作者的假说(尚未验证): 在同一平台上,持仓资金可以获得利息、或可用作保证金的合约,其长期热门的价格比不能获得利息的同类合约更接近真实结算率;两类合约在结算前一周的校准相同。什么结果会证明它错:两类合约的长期定价偏差相同,或可获利息的一类偏差更大。第八章给出检验锁仓折现本身的预注册方案。
小结
结算前不付利息的合约,把买家的资金锁到结算那天。在资金成本 下,真实概率为 的合约存在一条无套利带 ,,在带内任何价格都不给套利者足够的回报。带宽随期限与资金成本增长:对真实概率 0.99 的合约,一年期、14.5% 的资金成本下,带的下沿是 0.8646。冷门的高估与热门的低估都能在带内存活,所以期限越长,价格越偏向 0.5——同一个价格,一个月后与一年后代表不同的概率。
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).