关于「均值—方差前沿:三池配比、估计误差与收缩」,《达人矩阵的组合方差》给出的定量结论是什么?
结论:第五章说明短历史下的优化敌不过等额分配:等额的夏普比率 0.3638,用十二个季度优化只得 0.3033。
数据口径:书中借用组合理论的方差、相关、夏普比率与均值—方差前沿,只借其算术,讨论的是营销预算的分配,不是证券投资;本书不构成任何个人化的理财或投资建议,不推荐任何金融产品。 书中的达人、回报、相关、成本与预算均为示意;全部计算由书中程序生成;所有合作都须如实标注广告,不讨论购买流量或虚假评价。
对照数据表
下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.3033, 0.3638
| true Sharpe: 1/N 0.3638, oracle tangency 0.3941 | |||
| T=12 quarters (2000 runs) | |||
| plug-in | mean 0.3033 | p10 0.2587 | beats 1/N in 0.016 of runs |
| top 3 | mean 0.2936 | p10 0.2489 | beats 1/N in 0.000 of runs |
| T=40 quarters (2000 runs) | |||
| plug-in | mean 0.3273 | p10 0.2880 | beats 1/N in 0.075 of runs |
| top 3 | mean 0.3029 | p10 0.2635 | beats 1/N in 0.000 of runs |
查看完整验算程序
# Chapter 5. p52: optimising on a short history against plain 1/N.
# Model: 10 mid-tier creators with true means 0.18 to 0.288 (step 0.012),
# sd 0.80 each, every pair at rho 0.20, normal returns. A brand sees T
# quarters, then picks weights by: 1/N; plug-in tangency (sample means
# and sample covariance, negatives clipped to 0); "top 3" by sample
# mean; and the oracle tangency on true parameters. Each is scored by
# its TRUE Sharpe over the hurdle 0.08. 2,000 replications per T. With
# T <= 10 the sample covariance is singular and the plug-in rule cannot
# even be computed. Fixed seed; all parameters illustrative (示意).
import math
import random
import statistics
K, SD, RHO, HURDLE, REPS = 10, 0.80, 0.20, 0.08, 2000
MU = [0.18 + 0.012 * i for i in range(K)]
COV = [[SD * SD * (1.0 if i == j else RHO) for j in range(K)]
for i in range(K)]
rng = random.Random(20261113)
a, b = math.sqrt(RHO), math.sqrt(1.0 - RHO)
def solve(m, y):
n = len(y)
aug = [row[:] + [y[i]] for i, row in enumerate(m)]
for c in range(n):
p = max(range(c, n), key=lambda r: abs(aug[r][c]))
if abs(aug[p][c]) < 1e-12:
return None
aug[c], aug[p] = aug[p], aug[c]
for r in range(n):
if r != c:
f = aug[r][c] / aug[c][c]
aug[r] = [x - f * z for x, z in zip(aug[r], aug[c])]
return [aug[i][n] / aug[i][i] for i in range(n)]
def tangency(mu, cov):
z = solve(cov, [m - HURDLE for m in mu])
if z is None:
return None
z = [max(v, 0.0) for v in z]
tot = math.fsum(z)
return [v / tot for v in z] if tot > 0 else [1.0 / K] * K
def true_sharpe(w):
m = math.fsum(wi * mi for wi, mi in zip(w, MU))
v = math.fsum(w[i] * w[j] * COV[i][j] for i in range(K)
for j in range(K))
return (m - HURDLE) / math.sqrt(v)
equal = [1.0 / K] * K
oracle = tangency(MU, COV)
print("true Sharpe: 1/N %.4f, oracle tangency %.4f"
% (true_sharpe(equal), true_sharpe(oracle)))
for t in (12, 40):
res = {"plug-in": [], "top 3": []}
for _ in range(REPS):
hist = []
for _ in range(t):
f = rng.gauss(0.0, 1.0)
hist.append([MU[k] + SD * (a * f + b * rng.gauss(0.0, 1.0))
for k in range(K)])
cols = list(zip(*hist))
mu_hat = [statistics.fmean(c) for c in cols]
cov_hat = [[statistics.covariance(cols[i], cols[j])
for j in range(K)] for i in range(K)]
w = tangency(mu_hat, cov_hat)
res["plug-in"].append(true_sharpe(w if w else equal))
top = sorted(range(K), key=lambda k: -mu_hat[k])[:3]
res["top 3"].append(true_sharpe([1 / 3 if k in top else 0.0
for k in range(K)]))
print("T=%d quarters (%d runs)" % (t, REPS))
for name, xs in res.items():
s = sorted(xs)
print(" %-8s mean %.4f p10 %.4f beats 1/N in %.3f of runs"
% (name, statistics.fmean(xs), s[REPS // 10],
len([x for x in xs if x > true_sharpe(equal)]) / REPS))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《达人矩阵头部集中度与中腰部组合套利、尾部方差夏普比率与百名垂直达人投资组合母典》第 5 章「均值—方差前沿:三池配比、估计误差与收缩」(版本 1.0.0). 5D Library. DOI 待发布
DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-276 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026276,
author = {Fan, Yafu},
title = {Head Concentration versus Mid-Tier Portfolios in Creator Matrices: Tail Variance, Sharpe Ratios and the Allocation of a Budget across a Hundred Niche Creators},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}