Extreme-Value Theory and Generalised Pareto Tails: Whale-Order Ceilings and Reserves for Extreme Chargebacks: chapter answers
Shopkeepers, bookkeepers and those who run server rooms commonly draw the line for "unusually large" at the mean plus three standard deviations: orders above it are reviewed, reserves are held at it and server capacity is estimated from it. This book tests that ruler with extreme-value theory. For heavy-tailed data such as order values, chargeback losses and traffic bursts it systematically underestimates extremes, by a multiple set by the shape parameter of the tail that grows the deeper one goes, while the standard deviation itself is held hostage by a few whale orders. The book's orders, c…
DOI pending · monograph page goes live after publication · 5DL-B-244 · 中文版
- Chapter 1 · The Blind Spot of Mean Plus Three Standard Deviations: the One-in-a-Thousand Quantile and Whale Orders under Heavy Tails
Chapter 1 shows that on an illustrative order tail with shape about 0.4 the true one-in-a-thousand quantile is 2.57 times the normal estimate and the one-in-ten-thousand quantile 5.56 times; on samples the normal estima… - Chapter 2 · Peaks over Thresholds: the Pickands-Balkema-de Haan Theorem and the Convergence of Excesses
Chapter 2 presents the peaks-over-threshold theorem and shows that a lognormal looks like a power-law tail with shape 0.2148 at finite thresholds. - Chapter 3 · Probability-Weighted Moments: Estimating the Two Generalised Pareto Parameters with the Standard Library
Chapter 3 derives probability-weighted moments, whose root mean square error with 50 excesses is 0.169 against 0.209 for maximum likelihood. - Chapter 4 · Return Levels: How Large Is the Whale Order That Comes Once in So Many Months
Chapter 4 turns parameters into return levels: the exact once-in-ten-years whale order is 27214 yuan where the normal estimate is only 1459, and a once-in-ten-years event occurs within ten years with probability 0.632. - Chapter 5 · The Mean Excess Plot and the Choice of Threshold: Bias, Variance and the Lognormal Trap
Chapter 5 chooses the threshold with the mean excess plot and 150 simulated shops; the log error of the ten-year level is smallest at the ninetieth percentile, at 0.194. - Chapter 6 · Chargeback Reserves and Server Headroom: Writing Return Levels into the Ledger and the Server Room
Chapter 6 writes return levels into chargeback reserves and server headroom: at once in a thousand months the true monthly bill is 1.77 times the normal rule, a reserve spliced and fitted with the shape at its upper end… - Chapter 7 · Classical Sources East and West: "Treading on Frost, Hard Ice Comes" and the Extreme-Value Laws of Fisher and Tippett
Chapter 7 examines "treading on frost, hard ice comes" from Kun in the Book of Changes (on a heavy tail the chance of doubling once past a threshold is about one in five) and the extreme-value laws of Fisher and Tippett… - Chapter 8 · Closing the Empirical Loop: a Monthly Tail Check-Up, the Tail Reserve Estimator and Pre-Registered Tests
Chapter 8 gives a monthly tail check-up, the Tail Reserve Estimator and pre-registered tests: testing the normal one-in-a-thousand line needs only 280 orders, testing the peaks-over-threshold line 9100. Marked boxes sep…
Illustrative model calculation, not a live quote