关于「留存曲线是生存曲线:前三秒风险与完播门槛」,《录屏钩子赛马与收费站导流》给出的定量结论是什么?
结论:第三章用乘积极限估计读留存曲线,前三秒的离开占全部离开的 0.587,截尾把朴素完播率 0.299 修正为 0.318。
数据口径:两种打法各跑一年的合成实验里,对口的稳定案例录屏日均付款人数是赌爆款的 16.51 倍。 这是目标,不是承诺:在示意基线下,它需要每天 280,112 次播放。
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: 0.299, 0.318, 0.587
| A. Kaplan-Meier retention versus the true curve (2,000 plays, 35 s video) | |
| second KM S(t) true S(t) | |
| 1.00 | 0.836 0.835 |
| 2.00 | 0.702 0.698 |
| 3.00 | 0.603 0.583 |
| 5.00 | 0.557 0.538 |
| 10.00 | 0.450 0.440 |
| 20.00 | 0.390 0.379 |
| 30.00 | 0.342 0.326 |
| 34.99 | 0.318 0.303 |
| B. Exits 1,347, censored 653; exits in the first 3 s | 791 (0.587 of all exits) |
| naive completion (share of rows at the end) | 0.299; KM S(35) = 0.318 |
| C. Interval hazard per second from the KM curve | |
| 0.0-1.0 | 0.1794 |
| 1.0-2.0 | 0.1758 |
| 2.0-3.0 | 0.1533 |
| 3.0-5.0 | 0.0400 |
| 5.0-10.0 | 0.0425 |
| 10.0-20.0 | 0.0145 |
| 20.0-35.0 | 0.0135 |
查看完整验算程序
# Chapter 3. p31: a retention curve is a survival curve; estimate it with Kaplan-Meier.
# 2,000 synthetic plays of a 35-second video. Watch time follows a piecewise-constant hazard
# (high in seconds 0-3, lower after). Plays still running when the export was cut, or that
# reached the end, are censored, not counted as exits. Fixed seed; hazards are illustrative.
import math
import random
LENGTH = 35.0
HAZ = [(0.0, 3.0, 0.18), (3.0, 10.0, 0.04), (10.0, 35.0, 0.015)]
def draw(rng):
t = 0.0
for a, b, h in HAZ:
e = rng.expovariate(h)
if a + e < b:
return a + e, True
t = b
return t, False # watched to the end: censored at LENGTH
def km(data):
times = sorted(set(t for t, d in data if d))
s, out, at_risk = 1.0, {}, len(data)
data = sorted(data)
i = 0
for t in times:
while i < len(data) and data[i][0] < t:
at_risk = at_risk - 1
i = i + 1
d = len([1 for x, e in data if x == t and e])
s = s * (1.0 - d / at_risk)
out[t] = s
return out
def s_at(curve, t):
best = 1.0
for k in sorted(curve):
if k <= t:
best = curve[k]
return best
if __name__ == "__main__":
rng = random.Random(20261120)
data = []
for _ in range(2000):
t, d = draw(rng)
if rng.random() < 0.05: # export cut while the play was running
cut = rng.uniform(0.0, LENGTH)
if cut < t:
t, d = cut, False
t = round(t, 2)
data.append((t, d))
curve = km(data)
exact = lambda x: math.exp(-math.fsum([h * max(0.0, min(x, b) - a) for a, b, h in HAZ]))
print("A. Kaplan-Meier retention versus the true curve (2,000 plays, 35 s video)")
print(" second KM S(t) true S(t)")
for x in (1, 2, 3, 5, 10, 20, 30, 34.99):
print(" {:>6.2f} {:>9.3f} {:>11.3f}".format(x, s_at(curve, x), exact(x)))
exits = len([1 for t, d in data if d])
early = len([1 for t, d in data if d and t <= 3.0])
cens = len(data) - exits
print()
print("B. Exits {:,}, censored {:,}; exits in the first 3 s: {:,} ({:.3f} of all exits)".format(
exits, cens, early, early / exits))
naive = len([1 for t, d in data if t >= 34.99]) / len(data)
print(" naive completion (share of rows at the end) = {:.3f}; KM S(35) = {:.3f}".format(
naive, s_at(curve, 34.99)))
print()
print("C. Interval hazard per second from the KM curve")
for a, b in ((0, 1), (1, 2), (2, 3), (3, 5), (5, 10), (10, 20), (20, 34.99)):
sa, sb = s_at(curve, a - 1e-9) if a > 0 else 1.0, s_at(curve, b)
print(" {:>5.1f}-{:<5.1f} {:.4f}".format(a, b, -math.log(sb / sa) / (b - a)))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《零露脸案例录屏短视频量产、前三秒痛点钩子赛马与私域收费站导流母典》第 3 章「留存曲线是生存曲线:前三秒风险与完播门槛」(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-350 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026350,
author = {Fan, Yafu},
title = {Faceless Case-Screen Short-Video Production, Three-Second Hook Tournaments and Direct-to-Paywall Traffic Funnels},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}