5D Library › 短视频完播率风险函数、推荐反馈环与自动化获客漏斗 · English

第八章 十项审计与一项以视频为单位的预注册实验

十个问题

本章把全书写成十个可以由脚本检查的问题,适用于任何以短视频获客的业务,包括作者自己的业务。它们不是营销或经营建议。

  1. 中途再钩是否密于每 10 秒一次?(第一章)
  2. 是否在重复内容有两成以上会被识别时仍原样重发?(第二章)
  3. 是否以原始完播率或原始观看时长给视频排名?(第三章)
  4. 钩子宽度是否超过 0.5?(第四章)
  5. 是否在悬念结尾之后让观众一到达就看到出价?(第四章)
  6. 各账号是否共享设备或支付资料?(第五章)
  7. 现金储备是否不足 48 天的成本?(第五章)
  8. A/B 测试是否在第一个好看的日子就停止?(第六章)
  9. 是否只从头部账号复制做法?(第六章)
  10. 是否以发布后前三天的播放规划全年?(第六章)

审计程序

程序把十个问题写成十个自动标记,用于两个示意的运营:一个「朴素」运营(常见做法),一个按本书各章修正后的「审计后」运营。

推导

# Chapter 8. p15: Ten automatic flags for a short-video sales funnel.
# Each flag is a yes / no check on an operation's own settings; the
# chapter of this book that supplies the threshold is in brackets.
# Two illustrative operations: "naive" (common practice) and
# "audited" (after the corrections of this book).
FLAGS = [
    ("re-hooks closer than every 10 s [1]",
     lambda o: o["hook_gap_s"] < 10),
    ("duplicates reposted where 20 % or more are caught [2]",
     lambda o: o["reposts"] and o["dup_catch"] >= 0.2),
    ("videos ranked by raw completion or raw watch time [3]",
     lambda o: o["metric"] in ("completion", "watch")),
    ("hook breadth above 0.5 [4]",
     lambda o: o["breadth"] > 0.5),
    ("offer shown at arrival after an open story [4]",
     lambda o: o["offer_delay_s"] < 2 and o["open_story"]),
    ("accounts share devices or payment details [5]",
     lambda o: o["linked"]),
    ("cash reserve under 48 days of costs [5]",
     lambda o: o["reserve_days"] < 48),
    ("A/B tests stopped at the first good day [6]",
     lambda o: o["peeking"]),
    ("tactics copied from top accounts only [6]",
     lambda o: o["copy_top"]),
    ("year planned from the first three days of views [6]",
     lambda o: o["forecast_days"] <= 3),
]

OPS = {
    "naive": dict(hook_gap_s=4, reposts=True, dup_catch=0.5,
                  metric="completion", breadth=1.0, offer_delay_s=0,
                  open_story=True, linked=True, reserve_days=14,
                  peeking=True, copy_top=True, forecast_days=3),
    "audited": dict(hook_gap_s=10, reposts=False, dup_catch=0.5,
                    metric="normalised", breadth=0.314, offer_delay_s=8,
                    open_story=True, linked=False, reserve_days=60,
                    peeking=False, copy_top=False, forecast_days=60),
}

if __name__ == "__main__":
    for name, op in OPS.items():
        hits = [f for f, test in FLAGS if test(op)]
        print("   {}: {} of {} flags".format(name, len(hits), len(FLAGS)))
        for f in hits:
            print("     [x] " + f)
    print("\nReading: the naive operation trips every flag. The audited")
    print("one trips none, which says only that its settings agree with")
    print("this book's models; whether its buyers are better served is")
    print("a question for the experiment of p16.")
#    naive: 10 of 10 flags
#      [x] re-hooks closer than every 10 s [1]
#      [x] duplicates reposted where 20 % or more are caught [2]
#      [x] videos ranked by raw completion or raw watch time [3]
#      [x] hook breadth above 0.5 [4]
#      [x] offer shown at arrival after an open story [4]
#      [x] accounts share devices or payment details [5]
#      [x] cash reserve under 48 days of costs [5]
#      [x] A/B tests stopped at the first good day [6]
#      [x] tactics copied from top accounts only [6]
#      [x] year planned from the first three days of views [6]
#    audited: 0 of 10 flags
#
# Reading: the naive operation trips every flag. The audited
# one trips none, which says only that its settings agree with
# this book's models; whether its buyers are better served is
# a question for the experiment of p16.

朴素运营触发全部十个标记。审计后的运营一个也不触发,但这只说明它的设置与本书的模型一致,并不说明它的买家得到了更好的服务。后者只能由实验回答。

一项以视频为单位随机化的预注册实验

假说 · 尚未验证

实验方案。 每条新视频随机分配窄钩子(宽度 0.3)或宽钩子(宽度 1.0)。同一条视频的观众共享该视频的观众差异,因此分析单位是视频而不是观众。主要结局:每位观众的付费转化率,假设窄钩子为 0.010、宽钩子为 0.008;双侧 0.05,检验力 0.80;以视频为单位的 t 检验。由与作者无经济关系的独立分析者按预先登记的代码执行,结果不论正负都发表。

推导

# Chapter 8. p16: A pre-registered, video-randomised hook experiment.
# Each new video is randomly given a narrow hook (breadth 0.3) or a
# broad one (breadth 1.0). Viewers of one video share its audience
# draw, so the unit of analysis is the video, not the viewer.
# Primary outcome: paid conversion per viewer, assumed 0.010 under the
# narrow hook and 0.008 under the broad one at equal reach. Two-sided
# 0.05, power 0.80. Viewers per arm ignoring clustering:
#   n = (1.96 + 0.8416)^2 (p1 (1 - p1) + p2 (1 - p2)) / (p1 - p2)^2;
# with m viewers per video and intraclass correlation icc, multiply by
# the design effect 1 + (m - 1) icc and divide by m to get videos.
# A simulation of 2,000 null experiments (both arms at 0.010, icc
# 0.002, m = 2,000, 60 videos per arm) checks the type I error of a
# video-level t test.
import math
import random

Z = 1.96 + 0.8416

def viewers(p1, p2):
    return Z ** 2 * (p1 * (1 - p1) + p2 * (1 - p2)) / (p1 - p2) ** 2

if __name__ == "__main__":
    n = viewers(0.010, 0.008)
    print("   viewers per arm if viewers were independent: {:,.0f}".format(
        math.ceil(n)))
    print("   m viewers/video   icc     design effect   videos per arm")
    for m in (500, 2000, 8000):
        for icc in (0.0005, 0.002):
            de = 1 + (m - 1) * icc
            print("   {:15,d} {:7.4f} {:15.2f} {:16d}".format(
                m, icc, de, math.ceil(n * de / m)))
    rng = random.Random(16)
    sd_p = math.sqrt(0.002 * 0.010 * 0.990)
    rej = 0
    for _ in range(2000):
        arms = []
        for _ in range(2):
            rates = []
            for _ in range(60):
                p = min(max(rng.gauss(0.010, sd_p), 0.0), 1.0)
                k = rng.gauss(2000 * p, math.sqrt(2000 * p * (1 - p)))
                rates.append(k / 2000)
            arms.append(rates)
        ma, mb = (sum(a) / 60 for a in arms)
        va = sum((x - ma) ** 2 for x in arms[0]) / 59
        vb = sum((x - mb) ** 2 for x in arms[1]) / 59
        t = (ma - mb) / math.sqrt(va / 60 + vb / 60)
        rej += abs(t) > 1.98
    print("   null simulation, video-level t test: false alarms {:.4f}"
          .format(rej / 2000))
    print("\nReading: the experiment must count videos, not viewers:")
    print("between 22 and 140 videos per arm here, depending on how")
    print("much one video's audience differs from another's. Analysed")
    print("at the video level, the test keeps its stated error rate.")
#    viewers per arm if viewers were independent: 34,999
#    m viewers/video   icc     design effect   videos per arm
#                500  0.0005            1.25               88
#                500  0.0020            2.00              140
#              2,000  0.0005            2.00               35
#              2,000  0.0020            5.00               88
#              8,000  0.0005            5.00               22
#              8,000  0.0020           17.00               75
#    null simulation, video-level t test: false alarms 0.0465
#
# Reading: the experiment must count videos, not viewers:
# between 22 and 140 videos per arm here, depending on how
# much one video's audience differs from another's. Analysed
# at the video level, the test keeps its stated error rate.

若观众彼此独立,每组需要 34,999 名观众。但同一视频的观众并不独立:每条视频 2,000 名观众、组内相关 0.002 时,设计效应为 5.00,每组需要 88 条视频;组内相关 0.0005 时只需 35 条。每条视频 8,000 名观众时,两种组内相关分别需要 75 条与 22 条。对 2,000 次无效应实验的模拟中,以视频为单位的 t 检验误报率为 0.0465,与名义水平一致。

小结

十个问题(再钩密度、重发、指标、钩子宽度、出价时机、账号关联、现金储备、偷看、幸存者偏差、早期预测)都可以写成自动审计。检验钩子宽度的实验必须以视频为单位随机化与分析,所需的是几十到上百条视频,而不只是观众人数。

Fan, Y. (2026). Short-Video Completion Hazards, Recommender Exploration Loops and Automated Acquisition Funnels: Weibull Swipe-Away Kinetics, Multi-Stage Pool Promotion Thresholds and Traffic-to-Cashflow Conversion Bounds [短视频完播率风险函数、推荐反馈环与自动化获客漏斗] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090213 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).