Bayesian forgetting and Random Effects for forecasting TV ratings
summary
Claim: Bayesian filtering with a program-level random effects model and an estimated forgetting factor improves Channel 4 TVR forecasts versus fixed-forgetting RLS. Tools: mixed-effects regression, recursive RLS, conjugate normal‑inverse‑gamma updates and a simple particle filter to infer W. Evidence: 0.95 forgetting gives ~20 effective points so posteriors stay deliberately wide; adaptive W snaps to near zero on abrupt drops (Shipping Wars example) and yields lower error and more realistic uncertainty (about 21% vs 23% error). Channel 4, BARB, normal-inverse-gamma conjugacy, particle filter for W
AI-written summary · about
chapters (11)
- 0:00TV ratings forecasting intro
- 2:56Demographic targeting in TV ads
- 5:47Program ratings predictability
- 7:55Temporal and seasonal patterns
- 13:12Simple forgetting methods
- 14:38Bayesian forgetting with priors
- 21:39Theoretical justification for forgetting
- 24:53Optimal forgetting factors
- 27:19Variable forgetting in ratings
- 30:36Summary and references
- 31:19Q&A: Extending to other channels
related talks
30mIs your marketing effective? Let Bayes decide!
28mPredicting Retention on the BBC Sounds Service Using Historic User Dynamics
16mImproving Streaming Experience with Bayesian Optimization, from AB to...
25mScaling Bayesian Experimentation
30mKnowing what you don’t know matters- Uncertainty-aware model rating
21mLearning in situ: a randomized experiment in video streaming
42mSolving Real-World Business Problems with Bayesian Modeling
27mUsing Causal thinking to make Media Mix Modeling