meta-move-dist

Meta's Movement Distribution, kept beyond the 90 days HDX holds: for every administrative region of 217 countries and every day, how far from home the people of that region went, as the share at no distance, under 10 km, 10 to 100 km, and over 100 km. 2026-06-01 onward, from Facebook users who share their location.

CC BY 4.0. Credit Data for Good at Meta and the Humanitarian Data Exchange (see Licence).

Why this exists

AI for Good at Meta (formerly Data for Good at Meta) publishes Movement Distribution on the Humanitarian Data Exchange in files of four days to two weeks, and HDX keeps only about the last 90 days: older files are removed. A time series longer than three months exists only where someone kept the files. This dataset keeps every file it has fetched, unchanged, and rebuilds one table from all of them.

What is here

	rows	days
movement-distribution/movement_distribution_2026.parquet	11,735,260	2026-06-01 to 2026-08-31, 76 of 92

column
gadm_id, gadm_name	the region, its GADM id and name (JPN.41.51_1, Taitō)
country	ISO 3166-1 alpha-3
polygon_level	GADM level, an integer
home_to_ping_distance_category	0, (0, 10), [10, 100), 100+, km from home
distance_category_ping_fraction	the share of the region's people in that category that day
ds	the day, a date
source_file	added here: the original file the row was taken from

38,943 regions. Japan is 1,804 municipalities (GADM level 2), all 53 of Tokyo among them; 546,312 rows.

movement-distribution/original/ holds every file HDX has served, byte for byte: 12 CSVs and the readme PDF so far. manifest.json lists each with its HDX name and resource id, sizes, sha256, HDX's own dates, when it was fetched, and whether HDX still lists it.

Things to know

Days are missing. 16 days of August 2026 are in no file: 4, 5, 7, 9, 10, 13, 17, 19 to 22, 24, 26 and 28 to 30. The file named 2026-08-18_to_2026-08-31 holds 18, 23, 25, 27 and 31 only. Trust ds, not the file names.

Files overlap. The same days appear in two files at their edges (06-15 and 06-16, 07-01, 08-01; 617,668 rows), always with the same values. Each row is kept once, from the file HDX changed last. Four June files are cut by row count, not by day, so 06-19, 06-23 and 06-27 are split between two files.

The shares are noisy on purpose. Meta adds noise and drops small regions. 301,846 rows have a negative share, and the four shares of a region and day need not add up to 1.

Regions changed in March 2026. Meta re-drew some regions then; some that existed before do not now.

-- DuckDB, straight off the Hub: how far the people of Taito ward went
SELECT ds, home_to_ping_distance_category, round(distance_category_ping_fraction, 3)
FROM 'hf://datasets/yuiseki/meta-move-dist/movement-distribution/movement_distribution_2026.parquet'
WHERE country = 'JPN' AND gadm_id = 'JPN.41.51_1' AND ds = DATE '2026-07-14'
ORDER BY 2;

Keeping it

git clone https://github.com/yuiseki/meta-move-dist
cd meta-move-dist
uv run python scripts/01_download.py      # what HDX lists now; nothing is deleted
uv run python scripts/03_export_parquet.py
uv run python scripts/02_verify.py
uv run python src/publish.py --push

HDX adds a file every two weeks or so. A run that misses one for three months loses it for good.

Licence

CC BY 4.0, as HDX lists the dataset (cc-by). Credit as HDX asks:

Data for Good at Meta, Movement Distribution, Humanitarian Data Exchange
(https://data.humdata.org/dataset/movement-distribution)

The originals are Meta's, unchanged. The Parquet is made from them; LICENSE beside this file says what was done.