Kalman Filtering and State-Space Assimilation: Separating Observation Noise from True Signals in Wearable Readings and Advertising Reports: chapter answers
A glance at the wristband or the advertising report each day, followed by a decision for the day, rests on the common sense that newer is more accurate. This book tests that common sense with local level and local trend state space models: a reading is the true level plus measurement error, and when the error variance far exceeds the variance of the true daily change, yes-or-no decisions made on raw readings overreact to noise; with illustrative parameters raw readings flip the call 23.6 times per hundred days and the Kalman estimate only 5.5 times. After checking the sources, the book correc…
DOI pending · monograph page goes live after publication · 5DL-B-245 · 中文版
- Chapter 1 · Why Raw Readings Overreact: the Noise-to-Signal Ratio and Decision Flips
A reading is the true level plus measurement error. When the variance of the error far exceeds the variance of the true daily change, decisions taken on raw readings flip often, and most of those flips respond to error;… - Chapter 2 · Where the Kalman Gain Comes From: One Prediction, One Correction and Exponential Smoothing
The Kalman filter for the local level model is four lines of arithmetic: the predicted variance grows by $q$, the gain is the predicted variance over the predicted variance plus $r$, the estimate absorbs the innovation… - Chapter 3 · Trends and Steps: Slow Drift Takes Weeks to Show, a Real Jump Costs Days of Lag
Adding a slope state turns the local level model into the local trend model. With the illustrative parameters, a single reading's error is twelve times the daily drift, and comparing two readings tells almost nothing ab… - Chapter 4 · Looking Back and Filling Gaps: the RTS Smoother, Missing Days and Artefact Readings
In review, backward smoothing lowers the error of the middle days from 0.940 to 0.705, but it does nothing for the last day. Missing days need no imputation: the filter lets its uncertainty grow honestly and speeds up i… - Chapter 5 · How Large Is the Noise? Maximum Likelihood, Innovation Tests and Moments of Differences
Chapter 5 covers maximum likelihood, innovation tests and moments of differences: a gain estimated from four weeks of data lies between 0.031 and 0.481 and is only a starting point. - Chapter 6 · The Advertising Report: Attribution Noise, On/Off Flighting and Budget Switching
Chapter 6 turns to advertising reports: switching the budget on the raw report changes level 33.8 times in ninety days and earns 0.714 of the ideal, switching on the filtered estimate changes only 5.6 times and earns 0.… - Chapter 7 · Classical Sources East and West: the Mencius and Gauss's Least Squares
Chapter 7 examines the Mencius line "believing the book entirely is worse than having no book" and Gauss's least squares of 1809. - Chapter 8 · Closing the Loop: a Four-Week Filtered Log, Filtered Advertising Reports and Pre-Registered Tests
Chapter 8 gives the closed loops for a four-week filtered log and for filtered advertising reports, the Noisy-Reading Filter tool and two pre-registered protocols (96 advertising accounts under conservative assumptions,…
Illustrative model calculation, not a live quote