The cherry-picking machine

A statistics classroom built on real UMPD data, updated daily. See the code

"Crime is up 40%." Up 40% since when? A percent change needs a baseline, and whoever picks the baseline picks the story. This machine lets you do it on purpose — so you'll recognize it when a press release does it to you.

Build your own headline

Baseline period: to compared with: to

Pick two periods to generate a headline.

Monthly incident totals, with your baseline shaded blue and your comparison period shaded red. Try: baseline 2020 (the pandemic year) versus any normal year — instant "crime wave." Baseline 2016 versus 2020 — instant "historic decline." Same dataset, opposite stories.

The guardrails this site uses on itself

The Activity Log page auto-generates "story leads" like "Theft up 25% this July." That code refuses to print some comparisons. Its rules, straight from app.py:

Same window, both years. It compares July 1–12 this year against July 1–12 last year — never a full month against a partial one, and the same season against itself (crime has rhythms; September isn't January).
No small numbers. If either period has fewer than 5 incidents, no headline. Going from 1 robbery to 3 is "up 200%" — and meaningless. Small bases make percent changes explode.
No trivial changes. Anything under 10% is treated as ordinary fluctuation, not news.

For your story: before publishing any percent change, ask the three questions those rules encode: Is the comparison window fair? Are the raw counts big enough to matter? Is the change bigger than the data's normal wobble? (The spike detector makes that last question precise.)

Reproduce it in R

library(tidyverse)
library(lubridate)

activity <- read_csv("https://raw.githubusercontent.com/dwillis/umpd-logs/master/data/all-police-activity.csv")

monthly <- activity |>
  mutate(month = floor_date(as_date(`Date Occurred`), "month")) |>
  filter(month >= ymd("2015-01-01")) |>
  count(month)

# Change these four dates and watch the story flip.
baseline <- monthly |> filter(month >= ymd("2019-01-01"), month <= ymd("2019-12-01"))
compared <- monthly |> filter(month >= ymd("2021-01-01"), month <= ymd("2021-12-01"))

round((sum(compared$n) - sum(baseline$n)) / sum(baseline$n) * 100, 1)