ThGaskin/Migration_explorer_db

Migration explorer database

The database behind the Migration explorer, an interactive explorer for the UN migration flow estimates. For each country, it holds the estimates together with all the data used as training targets or inputs, the data that were considered and not used or excluded (with the reason), and where each value comes from.

Private, work in progress. The new estimates (source_id estimates, estimates_bilateral) come from the published ensemble fine-tuned on the updated training data. They are preliminary and not a release. The estimates of Gaskin & Abel (Nature 2026) are included for comparison as published_estimates and published_estimates_bilateral.

Contents

explorer.duckdb, the database, which each pipeline run replaces. Alongside it, in estimates/, the estimates of the fine-tuned ensemble as CSV files, in the format of the published estimates (also in the database):

• estimates/mig_unilateral.csv: by country and year, net migration (net), immigration (imm), emigration (emi), and the migrant stocks at the start of the year (imm_pop: foreign-born; emi_pop: born in the country, living abroad), as the ensemble mean, with the standard deviation in <column>_std
• estimates/mig_bilateral.csv: by origin, destination and year, the flows by previous residence (mig_prev) and by country of birth (mig_brth), and the bilateral stocks by country of birth at the start of the year (stock_mean), each with its standard deviation (mig_prev_std, mig_brth_std, stock_std)

The database is built by Explorer/build_db.py in the Migration_flows_extended repository, from the exports of its notebooks, the estimates, the UNPD candidate targets and the provenance descriptions in Explorer/sources.yaml. The explorer (Migration_explorer repository) downloads it at start and only displays it, without computing anything.

The table layout is defined in Explorer/schema.sql of the pipeline repository, which is the interface between the two repositories. The meta table records the build:

key	value
schema_version	version of the table layout (increased for changes that break the explorer)
built_at	time of the build
pipeline_commit	commit of Migration_flows_extended it was built from (+dirty: with uncommitted changes)

Tables

Table	Contents
countries	ISO 3166 alpha-3 code, name, alpha-2 code
sources	every source (source_id), with name, URL, description (markdown) and vintage
country_series	annual country series: net migration, immigration, emigration, and the migrant stocks foreign_born and born_abroad
net_migration_consistency	consistency of each national net migration series with the UN DESA stocks (the basis of its weight)
provenance	per country and quantity: what is used, what is excluded and why, dated decisions, caveats (markdown)
flows	bilateral flows by previous residence (residence) and by country of birth (birth): all flow data, the UNPD candidates, the estimates
flow_source_weights	the weights of the flow sources by source, reporter and direction
stocks	bilateral migrant stocks by country of birth: UN DESA revisions, corrected and start-of-year stocks, their changes between stock years (the targets), the estimates
notes	texts not tied to a country: source descriptions, flow weights, the stock methodology, the change history
covariates	country covariates (GDP, population, life expectancy, birth and death rates, conflict deaths, EU membership), with the source and kind of source of each value
bilateral_covariates	trade, refugees, distance, linguistic and religious similarity, colonial ties (non-zero values only)
data_vintages	the date of every download and of every input placed by hand

All values are in persons unless the schema says otherwise. Years are calendar years. Stocks and the stock-based series refer to the start of the year, and the UN DESA revisions to mid-year.

Status of a value

status	meaning
target	used as a training target, with its weight
reference	shown for comparison, not a target (e.g. WPP where national data exist)
not_used	considered but not used, e.g. a usable candidate that was not adopted
excluded	rejected, with the reason in reason
estimate	model estimate, with lower and upper (mean ± 1 s.d. of the ensemble)
input	covariates: the value used by the pipeline

Usage

Access needs a Hugging Face token with read access to this repository (hf auth login, or HF_TOKEN).

import duckdb
from huggingface_hub import hf_hub_download

path = hf_hub_download("ThGaskin/Migration_explorer_db", "explorer.duckdb", repo_type="dataset")
con = duckdb.connect(path, read_only=True)
con.sql("""
    SELECT year, source_id, value, status, weight
    FROM country_series
    WHERE iso = 'DEU' AND quantity = 'net_migration'
    ORDER BY year, source_id
""").df()

Sources

The data come from many providers: national statistical offices, UN DESA (WPP 2024, international migrant stocks), the UNPD flow database, Eurostat, QuantMig, UNHCR, the World Bank, IMF, UCDP, CEPII, USITC and others. The sources table lists each one with its URL and vintage, and their terms of use apply to the values taken from them. This database is a working tool for the project and is not for redistribution.