5D Library › Short-Video Completion Hazards, Recommender Exploration Loops and Automated Acquisition Funnels · 中文
Preface. Three Rules of Thumb for Short Video
Note
This book is not marketing, investment or business advice and does not describe any platform’s real algorithm. Please read this first. It studies the parts of short-video customer acquisition that arithmetic can test: in which second viewers swipe away, how a recommender’s tiered traffic pools promote videos and how many good videos they kill, why broader hooks bring more views and fewer sales, and the lags and risks between views and cash. Its Weibull parameters, pool sizes and thresholds, conversion rates, prices (in Malaysian ringgit, RM) and ban probabilities are illustrative model parameters, not data from any platform, account or business; the direction of each result comes from probability and statistics, the exact numbers move with the parameters.
People who acquire customers through short video widely believe three rules of thumb. First, the first three seconds decide everything and completion rate is the only metric that matters. Second, a platform first shows each new video to an initial pool of about 300 people, promotes the good ones step by step to pools of 3,000, 50,000 and on to a million, and a video that fails the first gate is a bad video. Third, more views are always better, and the broader and catchier the hook the better. This book turns the three rules into runnable programs and tests them one by one.
As in the other works of 5D Library, the book first tries to refute, then states the limits, and reports negative results as they are.
What the book contains
Chapter 1 describes swiping with a Weibull hazard, computes completion and mean watch time for videos of different lengths, and tests mid-video re-hooks. Chapter 2 simulates tiered traffic pools, computes how many good videos the first gate kills with a sample of 300, and compares reposting with re-editing as hedges. Chapter 3 shows the Matthew effect with a Pólya-type feedback model and asks which of completion rate, watch time and a length-normalised metric can be “gamed” by editing length. Chapter 4 finds the cash-optimal hook breadth and uses an attention-residue model to time the offer on a landing page. Chapter 5 simulates the cash flow from views to settled payments, and the risk of an account matrix under linked bans. Chapter 6 turns to evidence: sample size and peeking in hook A/B tests, survivorship bias from studying top accounts, view decay and early forecasts. Chapter 7 reads the Guiguzi, chapter “Chuai”, the Wenxin Diaolong, chapter “Zhangju”, and the Hanfeizi story of the man who bought the casket and returned the pearl. Chapter 8 gives a ten-flag audit and a pre-registered experiment randomised by video.
Claims corrected after checking
- “The first three seconds decide everything” broadly holds, but for a different reason than usually given. Under a falling hazard (Weibull shape 0.6), 0.320 of all swipes on a 60-second video happen in the first 3 seconds and 0.413 in the first 5; viewers who stay become ever less likely to leave, so completion falls with length far more slowly than an exponential model says (p01, p14).
- Failing the first gate does not make a video bad. With a sample of 300, even when every pool’s audience is alike, 0.169 of the top 10% of videos are stopped at the first gate; with audience differences between pools it is 0.308 to 0.478, in line with the folklore of “thirty to forty percent killed” (p03).
- Both completion rate and watch time can be raised by editing length. The first rewards short videos, the second long ones; only a comparison with the baseline for the same length, or a direct fit of holding power, restores the true order of quality (p06).
- Broader is not better. In this book’s illustrative model the broadest hook brings 10 times the views of the narrowest, yet earns only about a fifth of the daily net cash of the best breadth (p07).
- An account matrix is not diversification. Accounts that share devices or payment details are banned together: the chance of losing eight or more of ten rises from 0.0019 to 0.2353 (p10).
How to read the boxes
Boxes mark published evidence, runnable Python computations (16 programs, standard library only, fixed random seeds; each is run at build time and its real output printed below it), classical sources, notes and the author’s hypotheses; every hypothesis states what result would show it wrong. To keep code intact in print, the programs and their output use ASCII only, with terms in English as listed below.
Note
Code key: hazard; retention; completion (completion rate); mean watch (mean watch time); re-hook (a mid-video hook); pool (traffic pool); audience draw (difference between pool audiences); good / best videos (true completion in the top 10% / top 0.1%); repost; re-edit; Gini (Gini coefficient); holding power (Weibull scale L); breadth (hook breadth); buyers; net cash; rho (attention residue); settlement (payout by the payment platform); linked accounts; peeking (stopping a test early on a good day); survivorship (survivorship bias); icc (intraclass correlation); design effect; casket buyers (viewers drawn in by the hook who would never buy).
Disclosure
The author uses the works of 5D Library, including this book, as the methodological basis of his community courses, the 5D Ascension Cards (Dimensional Shift Card Decks) and the Yafu AI knowledge base; he runs a Southeast Asian influencer marketing and live-commerce business (AsiaInfluencerX), is building a two-sided platform of influencer profiles and merchant customer management (TalentDecX), and develops and runs paid online software and courses on traditional divinatory and mind-body themes. The models of this book apply directly to the short-video acquisition and conversion of these businesses, so he has a direct commercial interest in its subject. The book promotes no software, course, card deck or service, uses no customer, account or sales data from the author’s businesses, and accepts no sponsorship. All its models are offered for mathematical study and teaching only and constitute no marketing, investment or other business advice, nor a description of any platform’s real recommender. The experiment in Chapter 8 must be carried out with pre-registered code by an independent analyst who has no financial relationship with the author, and its results will be published whether positive or negative.
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).