Kaplan-Meier Retention Curves for Livestream Rooms: Locating Swipe-Away Hazard Spikes and Timing the Insertion of Hooks: chapter answers
The number most often used to review a livestream is average viewing time. This book refutes two beliefs behind it: that the average summarises how well a show keeps its audience, and that benefits or retention segments push the average up as long as they are spread evenly. Using the Kaplan-Meier product-limit estimator, the Nelson-Aalen cumulative hazard, piecewise-constant hazards and Bellman's dynamic programming, it shows that the risk of leaving a livestream room has spikes, one kind following each viewer's own moment of entry and another following segment switches on the programme clock…
DOI pending · monograph page goes live after publication · 5DL-B-277 · 中文版
- Chapter 1 · The Blind Spot of Average Viewing Time: Censoring and the Product Limit
Chapter 1 shows that censoring biases the average downwards: counting the end of the stream as leaving gives 408.1 seconds, counting only leavers gives 278.0, and the restricted mean from the product-limit estimator is… - Chapter 2 · From Retention to Hazard: the Nelson-Aalen Estimator and the Entry Spike
Chapter 2 locates the entry spike with the Nelson-Aalen estimator, the hazard in the first two seconds being 133.3 times that of settled viewers, and shows how casually marked spikes produce false reports. - Chapter 3 · Two Clocks: Seconds since Entry, the Programme Clock and Delayed Entry
Chapter 3 separates two clocks: the switch hazard ratio of settled viewers on the programme clock averages 8.52, squashed to 4.23 when newcomers are mixed in. - Chapter 4 · Piecewise-Constant Hazards: Maximum Likelihood, Likelihood Ratios and Finding the Switches
Chapter 4 compresses the switch spikes into two numbers with a piecewise-constant hazard, a hazard ratio of 9.55. - Chapter 5 · A Model of What a Hook Does: the Window, the Aftermath and Who Stays
Chapter 5 models what a hook does: a window covering the spike gains 693.4 audience-minutes, only 40.4 when kept viewers take and go, and those a hook keeps stay only 0.68 as long as those who would have stayed anyway. - Chapter 6 · Dynamic Programming for Hook Insertion: the Best Seconds for K Hooks
Chapter 6 schedules three hooks by dynamic programming for a gain of 1970.2, against 257.9 on average for one hook every 20 minutes; if the run sheet slips a minute and the hooks stay put, the gain falls to 210.7; sched… - Chapter 7 · Classical Sources East and West: Sunzi on Spirit and Graunt's Table of Mortality
Chapter 7 examines the Sunzi's "keen in the morning, slack at midday, homeward in the evening" and Graunt's mortality table of 1662 (hazard in the first six years 1.46 times the mid-life level). - Chapter 8 · Closing the Loop: a Hazard Review after Every Show and a Plan to Falsify
Chapter 8 gives the post-show review, the Swipe-Spike Hook Scheduler and a pre-registered test plan with an expected hazard ratio of 0.718. Marked boxes separate published evidence, runnable Python calculations (18 prog…
Illustrative model calculation, not a live quote