Skip to content

heavy_tails.md identifies power laws only visually — add a goodness-of-fit test #807

Description

@mmcky

heavy_tails.md defines power laws and Pareto tails formally, then identifies them by eye: "All plots are in log-log, so that a power law shows up as a linear log-log plot, at least in the upper tail." The lecture builds empirical CCDFs and Q-Q plots for firm size and city size, and stops there.

It has no goodness-of-fit test, no tail-index estimator, and no note that log-log linearity is a weak diagnostic. Searching the lecture for "goodness", "test for", "Hill estimator", "KS test", "Kolmogorov" or "Clauset" returns nothing.

Why this matters

Exercise ht_ex4 asks the reader to compare a Pareto distribution against a mean-and-median-matched lognormal, for the present discounted value of corporate tax revenue, and to observe the difference. The lecture therefore poses the Pareto-versus-lognormal question and gives the reader no way to settle it from data. The same comparison appears, also unresolved, as Exercise 2.2.10 of Economic Networks.

Eyeballing a log-log plot for straightness is precisely the practice the goodness-of-fit literature exists to caution against, so teaching only the visual method leaves readers with a diagnostic that looks more reliable than it is.

Suggested scope

A short section, not a new lecture:

  • Estimating the tail index — the Hill estimator, and its sensitivity to where the tail is deemed to start
  • Testing the hypothesis — the Clauset–Shalizi–Newman KS-based procedure is the standard reference and has a widely used implementation
  • A sentence of honesty in the existing visual material, noting that log-log linearity is suggestive rather than conclusive

Enough that a reader can answer "is this actually a power law?" instead of "does this look straight?". Wiring it back to ht_ex4 would close the loop on an exercise that currently ends in an observation rather than an answer.

Spun out of QuantEcon/meta#141, which collected it as the one concrete deliverable inside a broader proposal.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions