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

第八章 四步核查与一项预注册的锁仓检验

读一个预测市场价格之前的四个问题

本章把全书变成四个问题,可以向任何预测市场的价格——包括自己想交易的那一个——提出。它们不是投资建议,也不是交易策略。

一、这个价格是谁的哪个统计量? 价格是信念分布的某个分位数或加权均值,取决于交易者的预算与风险态度(第一章)。冷门的价格通常高于它的概率,热门的通常低于。问一问:交易者是谁,他们之间的分歧有多大。

二、距离结算还有多久,资金成本是多少? 结算前不付利息的合约存在一条无套利带,带宽随期限与资金成本增长(第三章,p06)。一份一年后结算的「确定」合约停在 0.92,并不意味着市场认为有 8% 的可能出错。

三、显示的套利能执行多少? 价格之和偏离 1、相关合约的价格不一致,都只是上限(第四、五章)。要算走盘口的成本、手续费、固定成本、腿部风险与锁仓期间的机会成本;还要读结算条款,确认「一套」真的必然支付 1。

四、关于偏差的证据有多少个独立事件? 一千份合约可能只来自二十个事件(第六章)。按事件聚类计算标准误;修正偏差之前,确认已结算的独立样本足够多,否则修正可能比偏差更有害。

一项预注册的锁仓检验

第三章的模型给出一个可以被数据否定的预测:在资金有成本的平台上,距离结算较远的热门合约会被系统性地低估。下面的检验在查看任何数据之前写成。

假说 · 尚未验证

研究方案。 样本:任何公开交易所存档中已结算的二元合约,取结算前恰好 90 天的价格,价格落在 [0.85, 0.97] 之内;每个事件最多取 5 份。主要估计量 DD:结算结果(0 或 1)减去价格的平均值;为正表示热门被低估。最小关注效应(SESOI):0.01。标准误按事件聚类。判定规则见程序;三道闸门使结果不可解释而非否定:找到的事件少于计划的 90%;超过 2% 的合约有结算争议;价格快照的时间规则在看到结果后被修改。本检验不推荐任何交易,由独立于任何交易决策的人员实施与分析。

推导

# Chapter 8. p16: A pre-registered test of the lock-up discount.
# Question: do long-dated favourites resolve YES more often than their
# price says? Sample: resolved binary contracts priced in [0.85, 0.97]
# exactly 90 days before resolution, five per event (the cap), from any
# public exchange archive. Estimand D = mean(outcome - price). The model
# of Chapter 3 predicts about +0.035; the smallest effect of interest
# (SESOI) is 0.01. Outcomes within an event share a shock (Gaussian
# copula, rho = 0.3), so standard errors are clustered by event.
# Part A: power at true D = +0.02 (half the model's prediction) and the
# false-positive rate at D = 0. Part B: the decision rule applied to one
# simulated archive.
import math
import random
from statistics import NormalDist

M, RHO, SESOI = 5, 0.3, 0.01
PHI = NormalDist()

def archive(rng, events, d):
    rows = []
    for e in range(events):
        z = rng.gauss(0, 1)
        for _ in range(M):
            price = rng.uniform(0.85, 0.97)
            p = min(price + d, 0.999)
            u = PHI.cdf(math.sqrt(RHO) * z + math.sqrt(1 - RHO)
                        * rng.gauss(0, 1))
            rows.append((e, price, 1 if u < p else 0))
    return rows

def estimate(rows):
    n = len(rows)
    d = sum(y - p for _, p, y in rows) / n
    by = {}
    for e, p, y in rows:
        by[e] = by.get(e, 0.0) + (y - p - d)
    se = math.sqrt(sum(v * v for v in by.values())) / n
    return d, se

def decide(d, se, planned, got, disputed):
    lo, hi = d - 1.96 * se, d + 1.96 * se
    if got < 0.9 * planned:
        return "UNINTERPRETABLE: fewer than 90 % of planned events"
    if disputed > 0.02:
        return "UNINTERPRETABLE: over 2 % of contracts disputed"
    if lo > 0 and d >= SESOI:
        return "SUPPORTED: favourites underpriced by at least the SESOI"
    if lo > -SESOI and hi < SESOI:
        return "EQUIVALENT: no lock-up discount of practical size"
    if hi < 0:
        return "REVERSED: favourites overpriced"
    return "INCONCLUSIVE: interval too wide to decide"

if __name__ == "__main__":
    rng = random.Random(59016)
    print("A. events   power at D=+0.02   false positives at D=0")
    for ev in (100, 200, 400, 600, 800):
        pw = fp = 0
        for _ in range(400):
            d, se = estimate(archive(rng, ev, 0.02))
            pw += (d - 1.96 * se > 0 and d >= SESOI)
            d, se = estimate(archive(rng, ev, 0.0))
            fp += (d - 1.96 * se > 0 and d >= SESOI)
        print("   {:6d}        {:5.3f}               {:5.3f}".format(
            ev, pw / 400, fp / 400))
    d, se = estimate(archive(rng, 600, 0.02))
    print("\nB. one archive of 600 events: D = {:+.4f}, 95 % CI"
          " [{:+.4f}, {:+.4f}]".format(d, d - 1.96 * se, d + 1.96 * se))
    print("   600 of 600 events found:")
    print("     " + decide(d, se, 600, 600, 0.0))
    print("   520 of 600 events found:")
    print("     " + decide(d, se, 600, 520, 0.0))
    print("\nReading: 600 events (3,000 contracts) give 0.95 power at")
    print("half the effect the model predicts; every branch of")
    print("the rule, including the ones that would embarrass the model,")
    print("is written down before any archive is opened.")
# A. events   power at D=+0.02   false positives at D=0
#       100        0.347               0.043
#       200        0.570               0.037
#       400        0.823               0.022
#       600        0.950               0.030
#       800        0.978               0.020
#
# B. one archive of 600 events: D = +0.0235, 95 % CI [+0.0123, +0.0347]
#    600 of 600 events found:
#      SUPPORTED: favourites underpriced by at least the SESOI
#    520 of 600 events found:
#      UNINTERPRETABLE: fewer than 90 % of planned events
#
# Reading: 600 events (3,000 contracts) give 0.95 power at
# half the effect the model predicts; every branch of
# the rule, including the ones that would embarrass the model,
# is written down before any archive is opened.

第三章的模型预测 DD 约为 +0.035。为了不让检验依赖于模型正确,把握度按一半的效应(+0.02)计算:100 个事件时只有 0.347,400 个时 0.823,600 个事件(3,000 份合约)时达到 0.950;D=0D = 0 时的假阳性率在各种规模下都低于 5%。所以计划样本是 600 个事件。

B 段在一份模拟存档上演示判定规则:D=+0.0235D = +0.0235,95% 置信区间 [+0.0123, +0.0347],下沿高于 0 且点估计不低于最小关注效应,判为「支持」。同一份存档若只找到 520 个事件,不足计划的 90%,判为「不可解释」——不论点估计是什么。规则为每个分支命名,包括会让本书的模型难堪的两个:区间落在 ±0.01 之内(「等价」:没有实际大小的锁仓折现),以及区间整体为负(「反向」:热门被高估)。

注

结果意味着什么、不意味着什么。 它描述的是所选平台、所选期限下的热门合约。正的 DD 与锁仓折现一致,但不排除其他解释——例如热门合约的买方恰好更厌恶风险;区分它们需要第三章假说中的对照:同一平台上可获利息与不可获利息的合约。DD 为正也不意味着买入这类合约有利可图:第三章已经说明,其年化回报未必胜过资金的机会成本。

本书没有做的事

本书没有使用任何真实交易所的数据,没有评估任何平台的合规性或可靠性,没有讨论特定司法辖区对预测市场的监管,也不建议读者参与任何预测市场或博彩活动。书中所有价格与盘口都是程序生成的合成数据。

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).