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
第五章 从播放到现金:滞后、厚尾与关联封号
一个有滞后的漏斗
播放不是现金。从一次播放到一笔到账的收入,中间隔着主页访问、私信、成交、平台结算,每一步都有比例损耗,也有时间滞后。程序设一个每天发布的账号:日播放服从中位数 4,000、对数标准差 1.0 的对数正态分布(大多数日子平淡,少数日子爆发);2% 的观众打开主页,其中 20% 发来私信,其中 15% 在 0 至 3 天后购买 RM 49 的入门产品;买家中 15% 在 14 天后购买 RM 298 的课程;平台在付款 7 天后结算。成本为第 0 天的 RM 3,000 启动费与每天 RM 120。
推导
# Chapter 5. p09: From views to cash: a funnel with lags and heavy tails.
# One account posts daily. Views per day are lognormal with median
# 4,000 and log-sd 1.0 (most days are quiet, a few go far). Funnel:
# 2 % of viewers open the profile; 20 % of those send a message;
# 15 % of those buy the RM 49 entry product, paid 0 to 3 days later;
# 15 % of buyers buy the RM 298 programme 14 days after entry.
# The payment platform settles each sale 7 days after it is paid.
# Costs: RM 3,000 set-up on day 0, then RM 120 per day. Counts are
# drawn as Poisson. 2,000 simulated 90-day runs, fixed seed.
import math
import random
def poisson(rng, lam):
if lam > 50:
return max(0, int(round(rng.gauss(lam, math.sqrt(lam)))))
k, p, l = 0, 1.0, math.exp(-lam)
while True:
p *= rng.random()
if p < l:
return k
k += 1
def run(rng, days=90):
cash = [0.0] * (days + 40)
cash[0] -= 3000
for d in range(days):
cash[d] -= 120
views = 4000 * math.exp(rng.gauss(0, 1.0))
msgs = poisson(rng, views * 0.02 * 0.20)
for _ in range(poisson(rng, msgs * 0.15)):
pay = d + rng.randint(0, 3)
cash[pay + 7] += 49
if rng.random() < 0.15:
cash[pay + 14 + 7] += 298
run_cash, out, be = 0.0, [], None
for d in range(days):
run_cash += cash[d]
out.append(run_cash)
if be is None and d > 0 and run_cash >= 0:
be = d
return out, be
if __name__ == "__main__":
rng = random.Random(49)
runs = [run(rng) for _ in range(2000)]
for day in (14, 30, 60, 89):
vals = sorted(r[0][day] for r in runs)
neg = sum(1 for v in vals if v < 0) / len(vals)
print(" day {:2d}: median cash RM {:9,.0f} P(cash < 0) {:.3f}"
.format(day, vals[len(vals) // 2], neg))
bes = sorted(r[1] if r[1] is not None else 999 for r in runs)
print(" break-even day: median {}, 10th pct {}, 90th pct {}".format(
bes[1000], bes[200], bes[1800]))
lam = 4000 * math.exp(0.5) * 0.02 * 0.20 * 0.15
print(" mean buyers per day {:.2f}; at the median day {:.2f}".format(
lam, 4000 * 0.02 * 0.20 * 0.15))
print("\nReading: settlements start only after a week, and the set-up")
print("cost is recovered around day 35, between day 26 and day 48 in")
print("eight runs out of ten. Heavy-tailed views make the mean")
print("day much better than the typical day, so plans built on an")
print("average day overstate what most months will bring.")
# day 14: median cash RM -3,673 P(cash < 0) 0.999
# day 30: median cash RM -1,081 P(cash < 0) 0.719
# day 60: median cash RM 6,342 P(cash < 0) 0.012
# day 89: median cash RM 13,686 P(cash < 0) 0.000
# break-even day: median 35, 10th pct 26, 90th pct 48
# mean buyers per day 3.96; at the median day 2.40
#
# Reading: settlements start only after a week, and the set-up
# cost is recovered around day 35, between day 26 and day 48 in
# eight runs out of ten. Heavy-tailed views make the mean
# day much better than the typical day, so plans built on an
# average day overstate what most months will bring.第 14 天现金的中位数为 RM 3,673 的赤字,几乎所有模拟都为负;第 30 天中位数仍是 RM 1,081 的赤字,0.719 的模拟为负;第 60 天中位数转为 RM 6,342 的盈余,只有 0.012 仍为负;第 89 天中位数为 RM 13,686。回本日的中位数是第 35 天,十次中有八次落在第 26 天至第 48 天之间。由于播放是厚尾的,平均每天有 3.96 位买家,而中位数那天只有 2.40 位:以「平均的一天」做计划,会高估大多数月份的收入。
注
操作上的含义。 现金储备至少要覆盖到回本日的较坏情形,即本模型中的第 48 天;结算周期、退款窗口与平台冻结都会把这一天推后。以中位数而不是平均数做月度计划。
账号矩阵与关联封号
同时运营多个账号常被当作分散风险。但分散只有在各账号的风险互相独立时才成立。程序设 10 个账号,每个账号每月被封的边际概率都是 0.03;封号可能来自共同冲击:某个月平台集中清理(概率 ),关联账号各以概率 被封,其余月份以较低的概率被封,使边际概率保持不变。
推导
# Chapter 5. p10: Ten accounts are not ten independent bets.
# A matrix of 10 accounts; each has the same marginal chance of a ban,
# 0.03 per month. Bans may come from a common shock: in a month with a
# platform sweep (chance s) each linked account is banned with chance
# q; otherwise with chance p0, where s q + (1 - s) p0 = 0.03.
# Scenarios: independent accounts; loosely linked (shared payment
# details); tightly linked (shared devices); and two separate clusters
# of five on different platforms, tightly linked within each. 12
# months, 40,000 runs each, fixed seed. Banned accounts stay banned.
import random
SCEN = [("independent", 0.0, 0.0),
("loosely linked", 0.05, 0.5),
("tightly linked", 0.04, 0.75)]
def months(rng, alive, s, q):
sweep = rng.random() < s
p0 = (0.03 - s * q) / (1 - s)
pr = q if sweep else p0
return [a and not (rng.random() < pr) for a in alive]
def run(rng, s, q, split=False):
if not split:
alive = [True] * 10
for _ in range(12):
alive = months(rng, alive, s, q)
return 10 - sum(alive)
a, b = [True] * 5, [True] * 5
for _ in range(12):
a = months(rng, a, s, q)
b = months(rng, b, s, q)
return 10 - sum(a) - sum(b)
if __name__ == "__main__":
rng = random.Random(10)
n = 40000
print(" scenario mean lost P(5 or more) P(8 or more)")
rows = [(name, s, q, False) for name, s, q in SCEN]
rows.append(("two clusters of five", 0.04, 0.75, True))
for name, s, q, split in rows:
lost = [run(rng, s, q, split) for _ in range(n)]
print(" {:22s} {:9.2f} {:13.4f} {:13.4f}".format(
name, sum(lost) / n, sum(1 for x in lost if x >= 5) / n,
sum(1 for x in lost if x >= 8) / n))
print("\nReading: the expected number of lost accounts is the same")
print("in every scenario, but linked accounts fail together. Shared")
print("devices or payment details turn a matrix into one large bet:")
print("the chance of losing eight or more rises from 0.0019 to")
print("0.2353. Two unlinked clusters cut that to 0.0966; only truly")
print("independent accounts restore the diversification.")
# scenario mean lost P(5 or more) P(8 or more)
# independent 3.07 0.1625 0.0019
# loosely linked 3.06 0.3521 0.1006
# tightly linked 3.02 0.3765 0.2353
# two clusters of five 3.06 0.3130 0.0966
#
# Reading: the expected number of lost accounts is the same
# in every scenario, but linked accounts fail together. Shared
# devices or payment details turn a matrix into one large bet:
# the chance of losing eight or more rises from 0.0019 to
# 0.2353. Two unlinked clusters cut that to 0.0966; only truly
# independent accounts restore the diversification.四种情形下,一年平均失去的账号数都约为 3 个(3.02 至 3.07)。差别全在尾部:账号独立时,一年失去八个以上的概率只有 0.0019;共享支付资料的松散关联时为 0.1006;共享设备的紧密关联时为 0.2353。把账号分成两个互不关联的五账号集群,这一概率降到 0.0966,但仍远高于完全独立。失去五个以上的概率也从 0.1625 升到 0.3765。关联把一个账号矩阵变成了一个大赌注。
假说 · 尚未验证
可检验的推论。 若封号主要来自共同冲击,则同一运营者旗下账号的封号时间应在时间上聚集,聚集程度随设备与支付资料的共享程度上升。什么会证明这一推论错:若封号时间在各账号之间近似独立,与共享程度无关,则矩阵确实分散了风险。
小结
漏斗的每一步都有损耗与滞后,现金在回本之前要先经历一段赤字;厚尾的播放使平均值误导计划。账号矩阵只有在各账号真正独立时才分散风险;共享设备或支付资料会让它们一起失败。
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).