关于「广告日报:归因噪音、开关试验与预算切换」,《可穿戴读数与广告日报的卡尔曼滤波》给出的定量结论是什么?

结论:第六章转到广告日报:按原始日报切换预算九十天切换 33.8 次、利润只有理想的 0.714,按滤波估计只切换 5.6 次、达到 0.903;花费不变时任何方法都学不到广告效果。

数据口径:本书用局部水平与局部趋势的状态空间模型检验这一常识:一个读数等于真实水平加测量误差,当误差方差远大于真实日变化方差时,按原始读数做是或否的决定会对噪音过度反应;示意参数下原始读数每百天翻转 23.6 次,卡尔曼滤波估计只翻转 5.5 次。 书中的噪音方差、漂移、回报率与预算均为示意参数;全部计算由书中程序生成。

English version

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: 0.714, 0.903, 33.8, 5.6

estimateswitches/90dwrong-day shareprofit (yuan)vs oracle
raw33.80.299111180.714
mean77.60.190137760.885
kalman5.60.175140600.903
oracle5.00.000155641.000
steady Kalman gain here: 0.1130
查看完整验算程序
# Chapter 6. p62: Switching the budget on raw vs filtered daily returns.
# True incremental margin per yuan of spend ("true return") starts at
# 1.05 and drifts as a random walk (sd 0.03 a day). The daily report
# shows true return + attribution noise (sd 0.25). Rule each morning:
# if the estimate of today's return is above 1, spend 2000 yuan, else
# 300 yuan (a floor that keeps the campaign measurable). Profit of a
# day = (true return - 1) x spend. Estimates: yesterday's raw report,
# trailing 7-day mean, Kalman (q = 0.03^2, r = 0.25^2), and the oracle
# that sees the true return. 1000 runs of 90 days, fixed seed.
import math
import random
import statistics

SD_Q, SD_R, DAYS, HI, LO = 0.03, 0.25, 90, 2000.0, 300.0
rng = random.Random(20261124)
res = {k: ([], [], []) for k in ("raw", "mean7", "kalman", "oracle")}
for _ in range(1000):
    rho, path, rep = 1.05, [], []
    for _ in range(DAYS + 1):
        path.append(rho)
        rep.append(rho + rng.gauss(0, SD_R))
        rho += rng.gauss(0, SD_Q)
    m, p = rep[0], SD_R ** 2
    kal = [m]
    for y in rep[1:]:
        pp = p + SD_Q ** 2
        k = pp / (pp + SD_R ** 2)
        m, p = m + k * (y - m), pp * (1 - k)
        kal.append(m)
    ests = {"raw": rep, "kalman": kal, "oracle": path[1:] + [0.0],
            "mean7": [statistics.fmean(rep[max(0, t - 6):t + 1])
                      for t in range(DAYS + 1)]}
    for name, e in ests.items():
        prof, flips, wrong, last = [], 0, 0, None
        for t in range(1, DAYS + 1):
            hi = e[t - 1] > 1.0
            prof.append((path[t] - 1) * (HI if hi else LO))
            wrong += hi != (path[t] > 1.0)
            if last is not None and hi != last:
                flips += 1
            last = hi
        res[name][0].append(math.fsum(prof))
        res[name][1].append(flips)
        res[name][2].append(wrong / DAYS)
top = statistics.fmean(res["oracle"][0])
print("estimate  switches/90d  wrong-day share  profit (yuan)  vs oracle")
for name in ("raw", "mean7", "kalman", "oracle"):
    pr = statistics.fmean(res[name][0])
    print("%-8s  %11.1f   %14.3f   %12.0f   %8.3f"
          % (name, statistics.fmean(res[name][1]),
             statistics.fmean(res[name][2]), pr, pr / top))
pss = (SD_Q ** 2 + math.sqrt(SD_Q ** 4 + 4 * SD_Q ** 2 * SD_R ** 2)) / 2
print("steady Kalman gain here: %.4f" % (pss / (pss + SD_R ** 2)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《卡尔曼滤波状态空间同化、可穿戴读数观测噪音与广告归因真实信号提纯母典》第 6 章「广告日报:归因噪音、开关试验与预算切换」(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-245 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026245,
  author    = {Fan, Yafu},
  title     = {Kalman Filtering and State-Space Assimilation: Separating Observation Noise from True Signals in Wearable Readings and Advertising Reports},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论